Pith. sign in

REVIEW 5 major objections 5 minor 65 references

Automatically Identify and Rectify: Robust Deep Contrastive Multi-view Clustering in Noisy Scenarios

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

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

Pith's one-line read Multi-view clustering can automatically identify and repair its own noisy samples, staying accurate when up to 90% of the data is corrupted.

desk verdict Solid empirical recipe, broken theory appendix, and an unverified clean-first-view assumption—worth a serious referee, not worth citing the theorem. read the letter →

arxiv 2505.21387 v1 pith:4GS3IGH2 submitted 2025-05-27 cs.CV

Xihong Yang , Siwei Wang , Fangdi Wang , Jiaqi Jin , Suyuan Liu , Yue Liu , En Zhu , Xinwang Liu
show 1 more author
Yueming Jin
This is my paper · ORCID
classification cs.CV
keywords multi-viewclusteringnoiseidentificationGaussianmixturemodelanomalydetectioncontrastivelearningrectificationunsupervisedrobustness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that multi-view clustering can resist real-world noise by treating noise as something to find and fix, not just to tolerate. The proposed framework, AIRMVC, flags each sample in each view as clean or noisy using a two-component Gaussian mixture model over a confidence score derived from the network's own soft predictions, then corrects flagged samples by mixing their class predictions toward a trusted first view. A contrastive loss whose sample pairs are filtered by prediction agreement supplies the representation, and Theorem 4.1 claims these representations provably keep clean information while discarding noisy information. On six benchmark datasets the method is reported to beat eleven prior algorithms on accuracy, NMI, and purity at every tested noise rate from 10% to 90%. If the claim holds, it matters because real multi-source data — sensor readings, news corpora, image features — routinely contains corrupted views, where existing methods degrade so badly they can fall below single-view baselines.

What carries the argument

The load-bearing object is the two-component Gaussian mixture model fitted to the soft-prediction confidence score $\chi_{y=q|i}$ (Eq. 6): one component, with higher mean, is interpreted as the clean cluster and yields $\varphi_i$, the other as the noisy cluster. It carries the argument because every downstream step keys off $\varphi_i$: the hybrid rectification $m^v_i = \varphi^v_i y^v_i + (1-\varphi^v_i) y^1_i$ (Eq. 7) interpolates between a view's own prediction and the first view's prediction with weight $(1-\varphi^v_i)$ on the latter, so samples judged noisy are rewritten toward the trusted view. The third mechanism is the thresholded contrastive pair filter, which admits a cross-view sample pair only when $(y^m_i)^\top (y^n_j) \geq \tau$, so soft-prediction agreement screens out incorrect positive/negative pairs before the contrastive loss acts. Theorem 4.1, proved via the data processing inequality and mutual-information bounds in the style of the referenced contrastive-learning analysis, converts these mechanisms into a claim about representation content: the learned $E^*$ provably keeps $y$ (clean information) and drops $y'$ (noisy information).

What would settle it

Two checks would settle the claim. First, reproduce the training recipe but inject a known fraction of noise into the designated first view as well: if accuracy and NMI fall as sharply as the baselines' do, the rectification anchor fails and the 'noise discarded' theorem does not transfer to the corrupt-anchor regime. Second, the authors should report the exact noise-injection protocol (which views were corrupted and with what noise distribution); if view 1 was never corrupted in the experiments, the assumption is satisfied by construction and the method's general robustness claim remains untested. A third, more direct test: with synthetic injection where ground-truth clean/noisy labels are known, check whether $\varphi_i$ actually flags the injected samples — if the GMM's clean probability is at chance level, the identification step, on which rectification and the theorem both depend, is not doing the claimed work.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that noisy multi-view clustering can be solved by turning noise identification into an anomaly-detection problem on soft predictions. AIRMVC builds a GMM over the representations, substitutes the network's soft prediction $p(y_i = k | x_i)$ for the GMM's discrete latent assignment, and derives a per-sample confidence score $\chi_{y=q|i}$, the probability that the predicted class matches the cluster implied by the representation. A two-component GMM fitted to this score yields $\varphi_i$, the probability that sample $i$ is clean, and $1-\varphi_i$, the probability that it is noisy. Rectification then blends each non-first view's prediction toward the first view's prediction, $m^v_i = \varphi^v_i y^v_i + (1-\varphi^v_i) y^1_i$, with a cross-entropy loss pulling the view's own prediction toward this mixture. The contrastive mechanism pairs samples across views only when their soft predictions agree above a threshold $\tau$, and Theorem 4.1 asserts the resulting representations $E^*$ satisfy $I(x;y) - \vartheta \leq I(E^*;y) \leq I(x;y)$ and $I(E^*;y') \leq I(x;y') - \eta + \vartheta$: clean information is retained, noisy information is shed. The identification is unsupervised and dynamic — the GMM parameters are updated alongside the network rather than fixed by a preprocessing pass — so the whole pipeline runs automatically.

Load-bearing premise

The entire rectification chain assumes the first view is clean (Section 3.2: 'we assume the first view to be the clean view'), because every other view's prediction is pulled toward the first view's prediction with weight $(1-\varphi_i)$; if view 1 itself carries noise, the correction step propagates and reinforces those wrong predictions instead of removing them. The paper never states how synthetic noise was injected in the experiments, so whether this assumption was actually satisfied — that is, whether view 1 was left uncorrupted — cannot be verified from the text.

Editorial extensions

If this is right

  • An unsupervised model can keep meaningful cluster structure up to 90% injected noise: on six benchmarks AIRMVC reports higher accuracy, NMI, and purity than eleven baselines at every tested noise rate from 10% to 90%.
  • Noise handling moves inside the training loop: the GMM's clean/noisy split updates together with the encoder, so identification, rectification, and representation learning reinforce one another instead of being separate stages.
  • Only one trusted view is needed: the rectification blend targets the first view's prediction, so the method requires no clean labels, no manual noise annotation, and no per-dataset tuning ($\alpha = \beta = 1.0$, $\tau = 0.8$ across all datasets).
  • Contrastive pairing becomes prediction-filtered: pairs are used only when soft predictions agree above $\tau$, which the paper argues removes most incorrect pairings, and Theorem 4.1 asserts these representations provably retain clean information while shedding noisy information.

Reading between the lines

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

  • The fix could transfer to single-view settings: Equation (7) is confidence-weighted interpolation toward a trusted prediction, the same soft-target mechanism used in label-noise and semi-supervised learning, so a known-clean anchor could denoise a single view the same way.
  • The paper leaves the reference-view choice fixed; a natural extension is to select the view with the highest average $\varphi$ as the anchor, or to blend toward all views weighted by their own clean probabilities, which would soften the dependence on one assumed-clean view.
  • Because performance rose monotonically as $\tau$ increased from 0.2 to 0.8, the method evidently prefers few, high-confidence pairs; a check at $\tau \to 1$, where only near-identical predictions pair, would reveal whether the contrastive signal remains useful at maximum precision.
  • Whether the claim generalizes to corrupt-anchor scenarios is untested: the paper never says how the synthetic noise was injected, so if view 1 was always spared, the reported results verify the method under its strongest assumption rather than under arbitrary noise patterns.
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

5 major / 5 minor

Summary. The paper proposes AIRMVC, a deep multi-view clustering method for noisy inputs. It identifies noisy samples by fitting a two-component GMM to a confidence score derived from per-view soft predictions, rectifies noisy predictions by mixing them with the first view's soft prediction, and trains with a soft-prediction-gated contrastive loss. The paper claims a theorem (Theorem 4.1) guaranteeing that the learned representations preserve clean information and discard noisy information, and validates the method on six benchmarks under 10%--90% injected noise against several recent baselines.

Significance. If the claims were fully established, the paper would be a useful contribution to robust multi-view clustering: it addresses a practical setting, the experiments are broad (six datasets, eleven baselines, three metrics, several noise levels), code is released, and ablations and hyperparameter analyses are included. The authors also explicitly state the clean-first-view assumption, which is a welcome transparency. However, the theoretical guarantee is currently not valid as presented, and the experimental protocol omits the information needed to check the key clean-view assumption. The contribution is therefore conditional on correcting the theory and clarifying or validating the noise setup.

major comments (5)
  1. [Appendix A.2, Eq. (16)] The proof of the first inequality in Theorem 4.1 invokes the data processing inequality for the Markov chain y ↔ x → E and states I(x;y) ≤ I(E;y). The data processing inequality gives the reverse inequality, I(E;y) ≤ I(x;y). Consequently, the derivation of the lower bound I(E*;y) ≥ I(x;y) − ϑ in Eq. (20) is unsupported. Because the same reversed inequality is used in the bound on I(E*;y′), the core claim that the learned representation discards noisy information is not established.
  2. [Appendix A.2, Eq. (17)] The equalities I(E*;x+) = I(x;x+) and I(E*;x+|y) = I(x;x+|y) are asserted without proof. Maximizing I(E;x+) over E does not imply that the data-processing upper bound is attained, and no argument is given for the conditional equality. These equalities are used to derive Eqs. (18)--(21), so the theorem is not proved even if the direction in Eq. (16) were corrected.
  3. [Section 4 and Appendix A.2] The theorem analyzes E* = argmax_E I(E; x+) for a generic positive sample x+; it never uses the confidence-gated indicator in Eq. (10), the threshold τ, the clean probability φ, or the rectification target in Eq. (7). The variables y and y′ are also not connected to the model's soft predictions or to the actual corruption process. The proof therefore does not justify the noise-robust or rectification properties that the paper attributes to the proposed mechanism.
  4. [Section 3.2, Eq. (7); Section 5.1] The hybrid rectification blends every view's soft prediction toward y^1 with weight (1−φ_i^v), and the text explicitly assumes the first view to be clean. However, Section 5.1 only states that noise is randomly introduced into the multi-view input data x at given proportions and does not specify whether view 1 is excluded from corruption or what noise model is used. If view 1 is corrupted for a sample, the rectification target y^1 is itself wrong and the loss in Eq. (8) propagates the error. The experiments need to report the noise injection protocol and either enforce the clean-view assumption or test its validity; otherwise the central empirical claim is not fully supported.
  5. [Section 3.1, Eqs. (5)--(8)] The identification probability φ_i is estimated from a two-component GMM fitted to χ_{y=q|i}, which is computed from the model's own soft predictions y_i, and those same soft predictions are then rectified using φ_i. In an unsupervised setting this self-referential loop may simply re-weight low-confidence samples rather than detect true corruption. To substantiate the claim of automatic identification, the authors should report a concrete validation, for example precision/recall of the identified noisy samples against the known injected noise, or a calibration check showing that φ_i tracks the actual noise status.
minor comments (5)
  1. [Section 5.1 and Table 1] The text lists seven datasets (BBCSport, Reuters, Caltech101, UCI-digit, WebKB, SUNRGB-D, and STL10) but Table 1 reports only six and does not include SUNRGB-D; the list should be reconciled.
  2. [Section 2 and Eq. (2)] The GMM is defined on representations E, but Eq. (2) writes p(q_i = k | x_i) ∝ N(x_i | μ_k, σ_k); the conditioning variable should be E_i for consistency with Eq. (3).
  3. [Appendix A.2] The proof is introduced as 'Theorem. 14', but the theorem is numbered 4.1 in the main text and A.1 in the appendix; the numbering should be consistent.
  4. [Section 5.4] There is a typo: 'we employt-SNE' should read 'we employ t-SNE'.
  5. [Section 6] The sentence 'Moreover, we theoretical proof that the learned representations could discard the noisy information' should be rephrased, e.g., 'we theoretically prove that the learned representations discard noisy information'.

Circularity Check

0 steps flagged · score 0.0 of 10

AIRMVC's empirical comparisons are independently benchmarked; the identified weaknesses (first-view-clean assumption, Theorem 4.1 proof gaps) are assumptions or proof errors, not circular reductions.

full rationale

No circular step can be exhibited by the paper's own equations. The noise-identification module (Eqs. 5-6) estimates a clean probability φ from the model's own soft-prediction confidences via a two-component GMM; this is a self-training / EM-style loop, but the reported clustering quality is evaluated against ground-truth labels on six external benchmarks, so the central empirical claim is not forced by the identification mechanism's internal definitions. The rectification step (Eq. 7) is an explicit convex combination of a view's soft prediction and the first view's prediction, and the paper openly states the load-bearing assumption 'we assume the first view to be the clean view,' attributing it to prior non-overlapping work (Huang et al., 2020; Yang et al., 2023a; Sun et al., 2024; Yang et al., 2021). An explicit assumption, even if unverified and potentially harmful when view 1 is noisy, is not a circular derivation. Theorem 4.1 has genuine proof problems: Eq. (16) states a reversed data-processing inequality, and the MI constraints in Definition A.4 are not linked to the GMM's φ or to the actual injected noise; these are correctness/validity gaps, not instances where the conclusion is identical to the input by construction. Self-citations (DealMVC, TGM-MVC, SCE-MVC) appear only as comparison baselines and related-work context and are not load-bearing for the method's derivation. The paper's main experimental claims are self-contained against external benchmarks, so the appropriate circularity finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 7 assumptions · 0 invented entities

The central mechanism rests on three fitted hyperparameters and several unverified modeling assumptions. The theoretical proof additionally rests on unproved equalities and a misapplied data-processing inequality. No new physical or mathematical entities are introduced.

free parameters (3)
  • tau (contrastive confidence threshold) = 0.8
    Chosen on all six datasets after sweeping {0.2, 0.4, 0.6, 0.8}; higher tau improves reported ACC/NMI/PUR, so the reported results use the best test-set value.
  • alpha (rectification loss weight) = 1.0
    Chosen from {0.01, 0.1, 1.0, 10} in sensitivity analysis using clustering metrics on the test sets; 1.0 is reported as optimal.
  • beta (contrastive loss weight) = 1.0
    Selected from the same sensitivity analysis as alpha; 1.0 is reported as optimal on test clustering metrics.
assumptions (7)
  • domain assumption The first view is clean.
    Section 3.2, Eq. (7): every other view's soft prediction is blended toward the first view's prediction, so this assumption is load-bearing and is not verified in the noise-injection experiments.
  • domain assumption Soft predictions can act as GMM latent assignments.
    Section 3.1 replaces p(q_i=k|x_i) with p(y_i=k|x_i) in Eq. (3), assuming the network's own predictions are reliable enough to define cluster means and variances in an unsupervised noisy setting.
  • domain assumption The confidence score chi is bimodally separable into clean and noisy components.
    Eq. (6) fits a two-component GMM to chi and calls the higher-mean component clean; no evidence of separability under the noise model is provided.
  • ad hoc to paper Minimizing the proposed contrastive loss maximizes I(E; x+).
    Definition A.3 in Appendix A.2 asserts this connection without proof and without relating Eq. (10) to an InfoNCE-style bound; Theorem 4.1 depends on it.
  • ad hoc to paper The maximizer E* satisfies I(E*; x+) = I(x; x+) and I(E*; x+|y) = I(x; x+|y).
    Eq. (17) in Appendix A.2 asserts these equalities to derive the theorem; they do not follow from the contrastive objective.
  • standard math Data Processing Inequality applies to y -> x -> E.
    Invoked in Appendix A.2, but Eq. (16) states the inequality in the wrong direction, giving I(x;y) <= I(E;y) instead of I(E;y) <= I(x;y).
  • domain assumption The mutual information constraints I(x; y|x+) <= vartheta and I(x; y'|x+) > eta hold.
    Definition A.4 assumes clean and noisy soft predictions are distinguishable by their conditional mutual information with the positive sample; this is a modeling assumption about the noise process, not something derived from the method.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automatically Identify and Rectify: Robust Deep Contrastive Multi-view Clustering in Noisy Scenarios." pith.science (2026). https://pith.science/paper/4GS3IGH2

@misc{pith2026250521387,
  author       = {Pith},
  title        = {Pith review of: Automatically Identify and Rectify: Robust Deep Contrastive Multi-view Clustering in Noisy Scenarios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4GS3IGH2}},
  note         = {Machine review of arXiv:2505.21387}
}
read the original abstract

Leveraging the powerful representation learning capabilities, deep multi-view clustering methods have demonstrated reliable performance by effectively integrating multi-source information from diverse views in recent years. Most existing methods rely on the assumption of clean views. However, noise is pervasive in real-world scenarios, leading to a significant degradation in performance. To tackle this problem, we propose a novel multi-view clustering framework for the automatic identification and rectification of noisy data, termed AIRMVC. Specifically, we reformulate noisy identification as an anomaly identification problem using GMM. We then design a hybrid rectification strategy to mitigate the adverse effects of noisy data based on the identification results. Furthermore, we introduce a noise-robust contrastive mechanism to generate reliable representations. Additionally, we provide a theoretical proof demonstrating that these representations can discard noisy information, thereby improving the performance of downstream tasks. Extensive experiments on six benchmark datasets demonstrate that AIRMVC outperforms state-of-the-art algorithms in terms of robustness in noisy scenarios. The code of AIRMVC are available at https://github.com/xihongyang1999/AIRMVC on Github.

Figures

Figures reproduced from arXiv: 2505.21387 by the authors.

Figure 1
Figure 1. An illustrative diagram of noise in a multi-view scenario. In the diagram, the areas marked with red exclamation points indicate instances where sensor failures or malfunctions at specific moments lead to data corruption. Compared to other views, these instances are considered noisy data. 1. Introduction In real-world scenarios, multi-source information data is prevalent. To effectively handle such data, Multi-View … view at source ↗
Figure 2
Figure 2. Illustration of the overall framework of the proposed AIRMVC. Specifically, we first encode the input multi-view data to generate representations. Next, an automatic noise identification and rectification strategy is introduced to mitigate the adverse impact of noisy data. Simultaneously, we propose a noise-robust contrastive mechanism to generate more reliable and discriminative representations for the downstream c… view at source ↗
Figure 3
Figure 3. Ablation studies for our proposed noisy-identification and rectification strategy on BBCSport dataset. automatically identify and rectify the noisy data, leading to a robust learning procedure. 2) When noise is present in multi-view data, the perfor￾mance of deep multi-view clustering models tends to be￾come unstable. The clustering performance of the model exhibits a downward trend as the noise proportion increases… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Ablation studies on BBCSport, Caltech101, STL10, and Reuters datasets with 10% noisy ratio. 20 Epoch 80 Epoch 140 Epoch 200 Epoch [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Visualization of the representations during the training process on UCI-digit dataset. each part of our design contributes to the overall clustering performance. 2) The removal of noisy identification and rectification strat￾egy (“(w/o) D&R”) results in a more substant…
Figure 6
Figure 6. Figure 6: Sensitive analysis on hyper-parameter τ on six datasets with 10% noisy ratio. BBCSport-ACC BBCSport-NMI WebKB-ACC WebKB-NMI Reuters-NMI Caltech101-NMI STL10-PUR UCI-digit-NMI [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Sensitivity Analysis for α and β with ACC, NMI, and PUR on BBCSport, WebKB, Reuters, UCI-digits, Caltech101 and STL10 datasets with 10% noisy ratio. To further examine the impact of the parameters α and β on our model, we performed sensitive experiments on six datasets…
Figure 8
Figure 8. Figure 8: Ablation studies on UCI-digit and WebKB datasets with 10% noisy ratio. • The presence of noise in multi-view data often destabilizes the performance of deep multi-view clustering models. As the proportion of noise increases, the clustering performance deteriorates. Thi…
Figure 9
Figure 9. Figure 9: Ablation studies on BBCSport, Caltech101, STL10, UCI-digit, WebKB and Reuters datasets with 30% noisy ratio. Sensitive analysis of threshold τ : We conduct experiments to evaluate the influence of the threshold parameter τ . We varied the value of τ within the range of…
Figure 10
Figure 10. Figure 10: Sensitivity Analysis for α and β with ACC, NMI, and PUR on BBCSport, WebKB, Reuters, Caltech101, UCI-digit, and STL10 datasets with 10% noisy ratio. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Visualization of the representations during the training process on UCI-digit dataset [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

65 extracted references · 53 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    A simple framework for contrastive learning of visual representations

    Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pp.\ 1597--1607. PMLR, 2020

    2020

  3. [3]

    Cover, T. M. Elements of information theory. John Wiley & Sons, 1999

    1999

  4. [4]

    Efficient and adaptive recommendation unlearning: A guided filtering framework to erase outdated preferences

    Dang, Y., Liu, Y., Yang, E., Guo, G., Jiang, L., Zhao, J., and Wang, X. Efficient and adaptive recommendation unlearning: A guided filtering framework to erase outdated preferences. ACM Transactions on Information Systems, 43 0 (2): 0 1--25, 2025 a

    work page 2025

  5. [5]

    Data augmentation as free lunch: Exploring the test-time augmentation for sequential recommendation

    Dang, Y., Liu, Y., Yang, E., Huang, M., Guo, G., Zhao, J., and Wang, X. Data augmentation as free lunch: Exploring the test-time augmentation for sequential recommendation. arXiv preprint arXiv:2504.04843, 2025 b

    arXiv 2025

  6. [6]

    Iterative deep structural graph contrast clustering for multiview raw data

    Dong, Z., Jin, J., Xiao, Y., Wang, S., Zhu, X., Liu, X., and Zhu, E. Iterative deep structural graph contrast clustering for multiview raw data. IEEE Transactions on Neural Networks and Learning Systems, pp.\ 1--13, 2023 a . doi:10.1109/TNNLS.2023.3313692

    work page arXiv 2023

  7. [7]

    Cross-view topology based consistent and complementary information for deep multi-view clustering

    Dong, Z., Wang, S., Jin, J., Liu, X., and Zhu, E. Cross-view topology based consistent and complementary information for deep multi-view clustering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 19440--19451, October 2023 b

    work page 2023

  8. [8]

    Robust contrastive multi-view clustering against dual noisy correspondence

    Guo, R., Yang, M., Lin, Y., Peng, X., and Hu, P. Robust contrastive multi-view clustering against dual noisy correspondence. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    work page

Show all 65 references
  1. [9]

    and Khasahmadi, A

    Hassani, K. and Khasahmadi, A. H. Contrastive multi-view representation learning on graphs. In International Conference on Machine Learning, pp.\ 4116--4126. PMLR, 2020

    2020

  2. [10]

    Momentum contrast for unsupervised visual representation learning

    He, K., Fan, H., Wu, Y., Xie, S., and Girshick, R. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 9729--9738, 2020

    2020

  3. [11]

    Hinton, G. E. Training products of experts by minimizing contrastive divergence. Neural computation, 14 0 (8): 0 1771--1800, 2002

    2002

  4. [12]

    D., Fedorov, A., Lavoie-Marchildon, S., Grewal, K., Bachman, P., Trischler, A., and Bengio, Y

    Hjelm, R. D., Fedorov, A., Lavoie-Marchildon, S., Grewal, K., Bachman, P., Trischler, A., and Bengio, Y. Learning deep representations by mutual information estimation and maximization. arXiv preprint arXiv:1808.06670, 2018

    2018 arXiv

  5. [13]

    Exploring the role of node diversity in directed graph representation learning

    Huang, J., Mo, Y., Hu, P., Shi, X., Yuan, S., Zhang, Z., and Zhu, X. Exploring the role of node diversity in directed graph representation learning. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pp.\ 2072--2080, 2024 a

    2024

  6. [14]

    On which nodes does gcn fail? enhancing gcn from the node perspective

    Huang, J., Shen, J., Shi, X., and Zhu, X. On which nodes does gcn fail? enhancing gcn from the node perspective. In Forty-first International Conference on Machine Learning, 2024 b

    2024

  7. [15]

    T., Lv, J., and Peng, X

    Huang, Z., Hu, P., Zhou, J. T., Lv, J., and Peng, X. Partially view-aligned clustering. Advances in Neural Information Processing Systems, 33: 0 2892--2902, 2020

    2020

  8. [16]

    and Dayan, P

    Hyv \"a rinen, A. and Dayan, P. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6 0 (4), 2005

    2005

  9. [17]

    B., and Kanagachidambaresan, G

    Imambi, S., Prakash, K. B., and Kanagachidambaresan, G. Pytorch. Programming with TensorFlow: solution for edge computing applications, pp.\ 87--104, 2021

    2021

  10. [18]

    Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  11. [19]

    Cross-view graph matching guided anchor alignment for incomplete multi-view clustering

    Li, X., Sun, Y., Sun, Q., Ren, Z., and Sun, Y. Cross-view graph matching guided anchor alignment for incomplete multi-view clustering. Information Fusion, 100: 0 101941, 2023

    2023

  12. [20]

    P., Sun, Y., Sun, Q., Sun, Y., W

    Li, X., Pan, Y. P., Sun, Y., Sun, Q., Sun, Y., W. Tsang, I., and Ren, Z. Incomplete multi-view clustering with paired and balanced dynamic anchor learning. IEEE Transactions on Multimedia, pp.\ 7087--7098, 2025

    2025

  13. [21]

    Consensus graph learning for multi-view clustering

    Li, Z., Tang, C., Liu, X., Zheng, X., Zhang, W., and Zhu, E. Consensus graph learning for multi-view clustering. IEEE Transactions on Multimedia, 24: 0 2461--2472, 2021

    2021

  14. [22]

    Efficient one-pass multi-view subspace clustering with consensus anchors

    Liu, S., Wang, S., Zhang, P., Xu, K., Liu, X., Zhang, C., and Gao, F. Efficient one-pass multi-view subspace clustering with consensus anchors. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp.\ 7576--7584, 2022 a

    2022

  15. [23]

    One pass late fusion multi-view clustering

    Liu, X., Liu, L., Liao, Q., Wang, S., Zhang, Y., Tu, W., Tang, C., Liu, J., and Zhu, E. One pass late fusion multi-view clustering. In International Conference on Machine Learning, pp.\ 6850--6859. PMLR, 2021

    2021

  16. [24]

    Deep graph clustering via dual correlation reduction

    Liu, Y., Tu, W., Zhou, S., Liu, X., Song, L., Yang, X., and Zhu, E. Deep graph clustering via dual correlation reduction. In AAAI Conference on Artificial Intelligence, 2022 b

    2022

  17. [25]

    Simple contrastive graph clustering

    Liu, Y., Yang, X., Zhou, S., Liu, X., Wang, S., Liang, K., Tu, W., and Li, L. Simple contrastive graph clustering. IEEE Transactions on Neural Networks and Learning Systems, 2023

    2023

  18. [26]

    Decoupled contrastive multi-view clustering with high-order random walks

    Lu, Y., Lin, Y., Yang, M., Peng, D., Hu, P., and Peng, X. Decoupled contrastive multi-view clustering with high-order random walks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 14193--14201, 2024

    2024

  19. [27]

    Revisiting self-supervised heterogeneous graph learning from spectral clustering perspective

    Mo, Y., Lu, Z., Yu, R., Zhu, X., and Wang, X. Revisiting self-supervised heterogeneous graph learning from spectral clustering perspective. In NeurIPS, 2024

    2024

  20. [28]

    Hg-adapter: Improving pre-trained heterogeneous graph neural networks with dual adapters

    Mo, Y., Yu, R., Zhu, X., and Wang, X. Hg-adapter: Improving pre-trained heterogeneous graph neural networks with dual adapters. In ICLR, 2025

    2025

  21. [29]

    Robust multi-view clustering with noisy correspondence

    Sun, Y., Qin, Y., Li, Y., Peng, D., Peng, X., and Hu, P. Robust multi-view clustering with noisy correspondence. IEEE Transactions on Knowledge and Data Engineering, 2024

    2024

  22. [30]

    Contrastive multiview coding

    Tian, Y., Krishnan, D., and Isola, P. Contrastive multiview coding. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part XI 16, pp.\ 776--794. Springer, 2020

    2020

  23. [31]

    J., Lokse, S., Jenssen, R., and Kampffmeyer, M

    Trosten, D. J., Lokse, S., Jenssen, R., and Kampffmeyer, M. Reconsidering representation alignment for multi-view clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 1255--1265, 2021

    2021

  24. [32]

    H., Wu, Y., Salakhutdinov, R., and Morency, L.-P

    Tsai, Y.-H. H., Wu, Y., Salakhutdinov, R., and Morency, L.-P. Self-supervised learning from a multi-view perspective. arXiv preprint arXiv:2006.05576, 2020

    2006 arXiv

  25. [33]

    Representation learning with contrastive predictive coding

    Van den Oord, A., Li, Y., Vinyals, O., et al. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2 0 (3): 0 4, 2018

    2018 arXiv

  26. [34]

    and Hinton, G

    Van der Maaten, L. and Hinton, G. Visualizing data using t-sne. Journal of machine learning research, 9 0 (11), 2008

    2008

  27. [35]

    Continual multi-view clustering

    Wan, X., Liu, J., Liang, W., Liu, X., Wen, Y., and Zhu, E. Continual multi-view clustering. In Proceedings of the 30th ACM International Conference on Multimedia, pp.\ 3676--3684, 2022

    2022

  28. [36]

    Fast continual multi-view clustering with incomplete views

    Wan, X., Xiao, B., Liu, X., Liu, J., Liang, W., and Zhu, E. Fast continual multi-view clustering with incomplete views. IEEE Transactions on Image Processing, 33: 0 2995--3008, 2024. doi:10.1109/TIP.2024.3388974

    2024

  29. [37]

    View gap matters: Cross-view topology and information decoupling for multi-view clustering

    Wang, F., Jin, J., Dong, Z., Yang, X., Feng, Y., Liu, X., Zhu, X., Wang, S., Liu, T., and Zhu, E. View gap matters: Cross-view topology and information decoupling for multi-view clustering. In Proceedings of the 32nd ACM International Conference on Multimedia, pp.\ 8431--8440, 2024 a

    2024

  30. [38]

    Evaluate then cooperate: Shapley-based view cooperation enhancement for multi-view clustering

    Wang, F., Jin, J., Hu, J., Liu, S., Yang, X., Wang, S., Liu, X., and Zhu, E. Evaluate then cooperate: Shapley-based view cooperation enhancement for multi-view clustering. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024 b

    2024

  31. [39]

    Generative partial multi-view clustering with adaptive fusion and cycle consistency

    Wang, Q., Ding, Z., Tao, Z., Gao, Q., and Fu, Y. Generative partial multi-view clustering with adaptive fusion and cycle consistency. IEEE Transactions on Image Processing, 30: 0 1771--1783, 2021

    2021

  32. [40]

    Incomplete multi-view clustering via graph regularized matrix factorization

    Wen, J., Zhang, Z., Xu, Y., and Zhong, Z. Incomplete multi-view clustering via graph regularized matrix factorization. In Proceedings of the European conference on computer vision (ECCV) workshops, pp.\ 0--0, 2018

    2018

  33. [41]

    Y., and He, L

    Xu, J., Ren, Y., Tang, H., Yang, Z., Pan, L., Yang, Y., Pu, X., Philip, S. Y., and He, L. Self-supervised discriminative feature learning for deep multi-view clustering. IEEE Transactions on Knowledge and Data Engineering, 35 0 (7): 0 7470--7482, 2022 a

    2022

  34. [42]

    Multi-level feature learning for contrastive multi-view clustering

    Xu, J., Tang, H., Ren, Y., Peng, L., Zhu, X., and He, L. Multi-level feature learning for contrastive multi-view clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 16051--16060, 2022 b

    2022

  35. [43]

    Investigating and mitigating the side effects of noisy views for self-supervised clustering algorithms in practical multi-view scenarios

    Xu, J., Ren, Y., Wang, X., Feng, L., Zhang, Z., Niu, G., and Zhu, X. Investigating and mitigating the side effects of noisy views for self-supervised clustering algorithms in practical multi-view scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Patte...

    2024

  36. [44]

    Partially view-aligned representation learning with noise-robust contrastive loss

    Yang, M., Li, Y., Huang, Z., Liu, Z., Hu, P., and Peng, X. Partially view-aligned representation learning with noise-robust contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 1134--1143, 2021

    2021

  37. [45]

    Robust multi-view clustering with incomplete information

    Yang, M., Li, Y., Hu, P., Bai, J., Lv, J., and Peng, X. Robust multi-view clustering with incomplete information. IEEE transactions on pattern analysis and machine intelligence, 45 0 (1): 0 1055--1069, 2023 a

    2023

  38. [46]

    Interpolation-based contrastive learning for few-label semi-supervised learning

    Yang, X., Hu, X., Zhou, S., Liu, X., and Zhu, E. Interpolation-based contrastive learning for few-label semi-supervised learning. IEEE Transactions on Neural Networks and Learning Systems, pp.\ 1--12, 2022. doi:10.1109/TNNLS.2022.3186512

    2022

  39. [47]

    Dealmvc: Dual contrastive calibration for multi-view clustering

    Yang, X., Jin, J., Wang, S., Liang, K., Liu, Y., Wen, Y., Liu, S., Zhou, S., Liu, X., and Zhu, E. Dealmvc: Dual contrastive calibration for multi-view clustering. In Proceedings of the 31th ACM International Conference on Multimedia, 2023 b

    2023

  40. [48]

    Cluster-guided contrastive graph clustering network

    Yang, X., Liu, Y., Zhou, S., Wang, S., Tu, W., Zheng, Q., Liu, X., Fang, L., and Zhu, E. Cluster-guided contrastive graph clustering network. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pp.\ 10834--10842, 2023 c

    2023

  41. [49]

    Z., Liu, X., and Zhu, E

    Yang, X., Tan, C., Liu, Y., Liang, K., Wang, S., Zhou, S., Xia, J., Li, S. Z., Liu, X., and Zhu, E. Convert: Contrastive graph clustering with reliable augmentation. In Proceedings of the 31st ACM International Conference on Multimedia, pp.\ 319--327, 2023 d

    2023

  42. [50]

    Hyperbolic contrastive learning for cross-domain recommendation

    Yang, X., Chang, H., Lai, Z., Yang, J., Li, X., Lu, Y., Wang, S., Yin, D., and Min, E. Hyperbolic contrastive learning for cross-domain recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pp.\ 2920--2929, 2024 a

    2024

  43. [51]

    Graphlearner: Graph node clustering with fully learnable augmentation

    Yang, X., Min, E., Liang, K., Liu, Y., Wang, S., Zhou, S., Wu, H., Liu, X., and Zhu, E. Graphlearner: Graph node clustering with fully learnable augmentation. In Proceedings of the 32nd ACM International Conference on Multimedia, pp.\ 5517--5526, 2024 b

    2024

  44. [52]

    Mixed graph contrastive network for semi-supervised node classification

    Yang, X., Wang, Y., Liu, Y., Wen, Y., Meng, L., Zhou, S., Liu, X., and Zhu, E. Mixed graph contrastive network for semi-supervised node classification. ACM Transactions on Knowledge Discovery from Data, 2024 c

    2024

  45. [53]

    Darec: A disentangled alignment framework for large language model and recommender system

    Yang, X., Jing, H., Zhang, Z., Wang, J., Niu, H., Wang, S., Lu, Y., Wang, J., Yin, D., Liu, X., Zhu, E., Lian, D., and Min, E. Darec: A disentangled alignment framework for large language model and recommender system. In 2025 IEEE 41rd International Conference on Data Engineer...

    2025

  46. [54]

    Dual test-time training for out-of-distribution recommender system

    Yang, X., Wang, Y., Chen, J., Fan, W., Zhao, X., Zhu, E., Liu, X., and Lian, D. Dual test-time training for out-of-distribution recommender system. IEEE Transactions on Knowledge and Data Engineering, 37 0 (6): 0 3312--3326, 2025 b . doi:10.1109/TKDE.2025.3548160

    2025

  47. [55]

    Apgl4sr: A generic framework with adaptive and personalized global collaborative information in sequential recommendation

    Yin, M., Wang, H., Xu, X., Wu, L., Zhao, S., Guo, W., Liu, Y., Tang, R., Lian, D., and Chen, E. Apgl4sr: A generic framework with adaptive and personalized global collaborative information in sequential recommendation. In Proceedings of the 32nd ACM international conference on...

    2023

  48. [56]

    Dataset regeneration for sequential recommendation

    Yin, M., Wang, H., Guo, W., Liu, Y., Zhang, S., Zhao, S., Lian, D., and Chen, E. Dataset regeneration for sequential recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.\ 3954--3965, 2024

    2024

  49. [57]

    Gzoo: Black-box node injection attack on graph neural networks via zeroth-order optimization

    Yu, H., Liang, K., Hu, D., Tu, W., Ma, C., Zhou, S., and Liu, X. Gzoo: Black-box node injection attack on graph neural networks via zeroth-order optimization. IEEE Transactions on Knowledge and Data Engineering, 2024 a

    2024

  50. [58]

    Dshield: Defending against backdoor attacks on graph neural networks via discrepancy learning

    Yu, H., Ma, C., Wan, X., Wang, J., Xiang, T., Shen, M., and Liu, X. Dshield: Defending against backdoor attacks on graph neural networks via discrepancy learning. Network and Distributed System Security Symposium, NDSS , 2025

    2025

  51. [59]

    Sparse low-rank multi-view subspace clustering with consensus anchors and unified bipartite graph

    Yu, S., Liu, S., Wang, S., et al. Sparse low-rank multi-view subspace clustering with consensus anchors and unified bipartite graph. IEEE Transactions on Neural Networks and Learning Systems, 2023 a

    2023

  52. [60]

    How to construct corresponding anchors for incomplete multiview clustering

    Yu, S., Wang, S., et al. How to construct corresponding anchors for incomplete multiview clustering. IEEE Transactions on Circuits and Systems for Video Technology, 34 0 (4): 0 2845--2860, 2023 b

    2023

  53. [61]

    Towards resource-friendly, extensible and stable incomplete multi-view clustering

    Yu, S., Dong, Z., Wang, S., et al. Towards resource-friendly, extensible and stable incomplete multi-view clustering. In International Conference on Machine Learning, pp.\ 57415--57440. PMLR, 2024 b

    2024

  54. [62]

    Cross-domain recommendation via user interest alignment

    Zhao, C., Zhao, H., He, M., Zhang, J., and Fan, J. Cross-domain recommendation via user interest alignment. In Proceedings of the ACM Web Conference 2023, pp.\ 887--896, 2023 a

    2023

  55. [63]

    Cross-domain recommendation via progressive structural alignment

    Zhao, C., Zhao, H., Li, X., He, M., Wang, J., and Fan, J. Cross-domain recommendation via progressive structural alignment. IEEE Transactions on Knowledge and Data Engineering, 2023 b

    2023

  56. [64]

    Asymmetric double-winged multi-view clustering network for exploring diverse and consistent information

    Zheng, Q., Yang, X., Wang, S., An, X., and Liu, Q. Asymmetric double-winged multi-view clustering network for exploring diverse and consistent information. Neural Networks, 179: 0 106563, 2024

    2024

  57. [65]

    Multiple kernel clustering with neighbor-kernel subspace segmentation

    Zhou, S., Liu, X., Li, M., Zhu, E., Liu, L., Zhang, C., and Yin, J. Multiple kernel clustering with neighbor-kernel subspace segmentation. IEEE transactions on neural networks and learning systems, 31 0 (4): 0 1351--1362, 2019

    2019

Pith tools

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