Pith. sign in

REVIEW 4 major objections 6 minor 51 references

DWCL: Dual-Weighted Contrastive Learning for Multi-View Clustering

T0 review · 4 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Multi-view clustering anchors all contrastive pairs to one best view, beating pairwise baselines on eight datasets while cutting cost from O(V²) to O(V).

desk verdict A solid empirical multi-view clustering recipe with a defensible new contrastive mechanism, packaged with a theory section that should be ignored. read the letter →

arxiv 2411.17354 v2 pith:AAXTV375 submitted 2024-11-26 cs.CV cs.LG

Hanning Yuan , Zhihui Zhang , Qi Guo , Lianhua Chi , Sijie Ruan , Jinhui Pang , Xiaoshuai Hao This is my paper · ORCID
classification cs.CVcs.LG
keywords multi-viewclusteringcontrastivelearningrepresentationdegenerationsilhouettecoefficientBest-Othermechanismviewqualityweightdiscrepancydeep
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 the usual way of building contrastive pairs in multi-view clustering—combining every pair of views—produces many unreliable cross-views and lets low-quality views pull down the shared representation. The authors propose a Best-Other (B-O) mechanism that pairs each view only with the view that has the highest silhouette coefficient, and a dual weighting that downweights cross-views that are both low in quality and high in discrepancy. They argue this mitigates representation degeneration, the phenomenon where contrastive learning makes views too similar and destroys individual-view information. If correct, the method would make multi-view contrastive clustering more accurate, more robust to weak views, and markedly cheaper to run.

What carries the argument

The central object is the Best-Other cross-view, defined as the pair formed by the view with the highest silhouette coefficient (SI), computed from k-means cluster assignments on each view's representation, and each other view in turn. The silhouette coefficient acts as a view-quality proxy, producing an exponential quality weight $W_{SI}=e^{SI}$; this is multiplied by the view-discrepancy weight $W_{CMI}=e^{CMI}-1$ from prior work, which downweights cross-views whose cluster-label distributions disagree strongly. The resulting dual weight $W_{Dual}=W_{CMI}\cdot W_{SI}$ is applied to an InfoNCE contrastive loss, and the best view and weights are alternately updated during fine-tuning.

What would settle it

Find or construct a multi-view dataset where the view with the highest silhouette coefficient has lower true clustering quality (measured by agreement with ground-truth labels) than another view, then run DWCL against the pairwise mechanism: if pairwise achieves higher clustering accuracy or NMI, the central claim that the B-O mechanism anchored by SI reliably improves clustering fails.

Watch

Extended reading notes

Core claim

The central claim is that replacing pairwise cross-view construction with a Best-Other contrastive mechanism, guided by the silhouette coefficient, and weighting each cross-view by the product of a view-quality weight and a view-discrepancy weight, prevents representation degeneration and improves clustering. The paper reports that DWCL outperforms nine prior multi-view contrastive clustering methods and the best-single-view baseline across eight datasets, with absolute accuracy gains of about 3.5% on Caltech5V7 and 4.4% on CIFAR10 over the second-best method, and consistent gains on all views within each dataset. The authors also prove that their dual-weighted loss provides a better lower bound on mutual information than the self-weighted approach and that the B-O mechanism cuts the number of contrastive cross-views from quadratic to linear.

Load-bearing premise

The load-bearing premise is that the silhouette coefficient computed from k-means on the current representation reliably identifies the view that is best to anchor all contrastive pairs; if the highest-SI view is not the most informative one, the entire contrastive mechanism is anchored to a suboptimal view.

Editorial extensions

If this is right

  • If the highest-silhouette view is a reliable quality anchor, multi-view contrastive clustering can be made both faster and more accurate by discarding all non-anchor pairwise cross-views.
  • The dual weighting should suppress low-quality, high-discrepancy cross-views that otherwise dominate the loss and cause representation degeneration.
  • Per-view representations should improve uniformly rather than only for the best view, because every view is contrasted against the same high-quality anchor.
  • Computational cost drops from $O(|V|^2)$ to $O(|V|)$ cross-views, making the method more practical for datasets with many views.
  • The learned representations transfer to downstream tasks such as linear classification, improving accuracy, recall, and F1 on tested datasets.

Reading between the lines

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

  • A testable extension is to replace the silhouette coefficient with another view-quality estimator (e.g., clustering stability or mutual information with a consensus partition) and check whether performance degrades when the SI-selected view differs from the one chosen by the alternative estimator.
  • The method implicitly assumes that view quality is a property of the view itself, but quality may be instance-dependent; a view that is globally low-SI could still contain reliable information for specific clusters, which the B-O mechanism would discard.
  • The abstract in the paper reports 5.4% and 5.6% absolute gains on Caltech6V7 and MSRCv1, while the introduction and experiments report 3.5% and 4.4% on Caltech5V7 and CIFAR10; a reader comparing numbers should verify which datasets and gains the final version intends.
  • Because the best view is re-selected during training, the anchor can change between iterations, which may cause instability on datasets where two views have near-equal silhouette scores; this is an empirical risk worth probing.
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

4 major / 6 minor

Summary. The paper proposes DWCL, a multi-view contrastive clustering method. It selects a "best" view using the silhouette coefficient (SI), constructs contrastive pairs only between that view and each other view (the Best-Other, or B-O, mechanism), and weights the contrastive loss by the product of a view-quality weight e^{SI} and a view-discrepancy weight derived from normalized mutual information between cluster-label distributions. The authors report consistent improvements over nine baselines across eight multi-view datasets, ablation studies supporting both the B-O mechanism and the dual weighting, an efficiency comparison, and two theorems intended to justify the mechanism theoretically.

Significance. If the empirical results hold, DWCL is a useful and efficient contribution: the B-O mechanism reduces cross-view pairs from O(V^2) to O(V), and the reported 10-run mean±std results consistently favor DWCL over the compared baselines, with ablations showing that both the B-O mechanism and the dual weights contribute. The efficiency comparison is also a positive feature. However, the theoretical validation currently claimed in the paper is not sound: Theorem 1 conflates the InfoNCE lower bound with mutual information equality, and the "better lower bound" argument reduces to multiplying a bound by constants; Theorem 2 restates the algorithmic construction rather than proving the superiority of the SI-based anchor choice. Because the authors list theoretical justification as a contribution, these parts must be corrected or substantially reframed before the paper's claims are credible.

major comments (4)
  1. [Section 3.4, Theorem 1 (Eqs. 22-23)] The claimed equivalence is not established. The proof cites Oord et al. to assert that minimizing the InfoNCE loss equals maximizing mutual information, but the cited result establishes that InfoNCE is a lower bound on mutual information, not an equality. Moreover, σ is defined in the theorem as the mutual information between high-level representations, while the proof derives the bound on W_CMI using entropies of cluster-label distributions in Eq. (17), so the inequality W_CMI ≥ e^{σ/log N}−1 does not follow from the stated assumptions. The step from Eq. (22) to Eq. (23) therefore multiplies a lower bound by e^{α+β}>1 and calls the result a better lower bound; this is an algebraic restatement, not a proof of improved optimization. In addition, the claim that SI(ĥ_i)∈(0,1) for all points under k-means is not generally true, because silhouette values can be negative, so W_SI∈(1,e) is not guaranteed.
  2. [Section 3.4, Theorem 2] The theorem's conclusion that B-O "prevents low-quality cross-views from being augmented" is a direct restatement of the algorithm: by construction, all non-best pairs receive zero weight. The proof establishes only a complexity reduction from O(V^2) to O(V) and a range comparison for the weights; it does not show that selecting the anchor by the highest silhouette coefficient yields a superior or even reliable contrastive objective. Thus the theoretical analysis does not validate the central design choice; it formalizes the mechanism rather than proving its effectiveness.
  3. [Section 3.2-3.3, Table 7] The method's view-quality measure is the silhouette coefficient, but the paper's own per-view results undermine this premise. On NUSWIDE, View 2 has the second-highest SI (1.0×10) yet the lowest initial clustering accuracy (30.0), while View 4 has among the lowest SI values (0.6×10) yet the second-highest initial accuracy (38.7). Consequently, the weight e^{SI} up-weights a lower-accuracy view relative to a higher-accuracy one, and since the best view is chosen by argmax SI, all V contrastive pairs are anchored to a potentially suboptimal view. The conclusion explicitly concedes that "the best view may not always be the ideal choice in every scenario." The authors should either provide evidence that SI ranking correlates with actual view quality across training iterations and datasets, or replace the criterion with a more robust view-quality measure.
  4. [Section 3.4 and Section 5] The paper claims "rigorous theoretical justifications" for both efficiency and effectiveness, but the two theorems as written do not deliver this. Theorem 1 is a tautological scaling of a lower bound, and Theorem 2 is a complexity count plus a definitional zeroing of non-best pairs. The authors should either remove the theoretical claims or substantially rewrite them so that they make falsifiable statements about representation quality, rather than restating the weighting scheme.
minor comments (6)
  1. [Abstract] The abstract in the header reports gains on Caltech6V7 and MSRCv1 of 5.4% and 5.6%, while the full-text abstract reports gains on Caltech5V7 and CIFAR10 of 3.5% and 4.4%; these versions must be reconciled.
  2. [Eqs. (14)-(15)] The hyper-parameter λ appears both as the multiplier of Σ_v L_v^R in Eq. (14) and inside the definition of L_v^R in Eq. (15), which double-counts the same weight; one of the two appearances should be removed.
  3. [Eq. (13)] The text says H(ŷ_v) is the "cross-entropy" of the label distribution, but the expression is the Shannon entropy; the terminology should be corrected.
  4. [Table 8 and Section 4.3] The notation "W_SIL" is used in several places instead of the defined "W_SI"; the subscripts should be unified throughout.
  5. [Figure 3 caption] The caption says "seven baseline methods and BSV," but the comparison includes nine named baselines plus BSV; the count should be corrected.
  6. [Algorithm 1, line 3] There is a typo: "Opitimize" should be "Optimize."

Circularity Check

2 steps flagged · score 5.0 of 10

Theoretical validation is definitional: Theorem 1's 'better lower bound' is just the inserted e^{SI} multiplier, and Theorem 2's efficiency restates the B-O definition; the empirical benchmarking is independent.

  1. self definitional [Section 3.4, Theorem 1, Eqs. (10)-(11), (21)-(23)]
    "The view quality weights W^{v_i}_{SI} and W^{v_j}_{SI} are expressed as W^{v_i}_{SI}=e^α∈(1,e), W^{v_j}_{SI}=e^β∈(1,e). Thus, minimizing the dual weighted contrastive loss ... is equivalent to maximizing e^{α+β}(e^{σ/logN}−1)I(Ĥ^{v_i},Ĥ^{v_j}), where e^{α+β}>1. Therefore, DWCL has better lower bounds than SEM."

    The 'better lower bound' is manufactured by the paper's own definition: the quality weight W_SI=e^{SI} is inserted into the loss in Eq. (10)-(11), so the extra factor e^{α+β}>1 appears by construction. Comparing the weighted loss to SEM's bound is therefore comparing SEM's loss to SEM's loss multiplied by a constant greater than one; the claimed improvement is exactly the inserted multiplier. No independent argument shows that a larger constant on the same mutual-information expression yields a tighter or more achievable bound.

  2. self definitional [Section 3.4, Theorem 2, Eqs. (24)-(27)]
    "In the B-O contrastive mechanism, if SI(Ĥ^{v_i}) or SI(Ĥ^{v_j}) is not maximum in {SI(Ĥ^1),SI(Ĥ^2),...,SI(Ĥ^V)}, letting W^{v_i,v_j}_{SI}=0, i.e. the total contrastive loss is calculated as ... where the quantity of cross-views is just V, the complexity is reduced from quadratic to linear, i.e., from O(|V|^2) to O(|V|), decreasing the computational costs."

    The complexity reduction is the definition of the B-O mechanism, not a derived result: B-O is defined as pairing the highest-SI view with each other view and zeroing all other pairs. Counting the resulting V pairs and calling this a complexity improvement simply restates the construction. Similarly, 'preventing low-quality cross-views from being augmented' is the direct effect of setting their weights to zero, which is exactly how B-O was defined. Thus Theorem 2 validates the method by unpacking its own definition.

full rationale

The paper's central empirical claim, that DWCL outperforms nine baselines across eight multi-view datasets, is evaluated against external methods and is not circular; those comparisons (Tabs. 5-11) stand independently of the theoretical section. The circularity is confined to the theoretical framing. Theorem 1's conclusion that DWCL has better lower bounds than SEM is true by construction: the dual weight defined in Eqs. (10)-(11) multiplies the SEM-style loss by e^{α+β}>1, so the 'better bound' is literally the inserted factor. Theorem 2's efficiency claim is also definitional: the B-O mechanism is defined as keeping only the V pairs involving the highest-SI view, so the O(V) count and the suppression of other pairs are immediate consequences of the definition rather than validated properties. There is no load-bearing self-citation and no fitted parameter renamed as a prediction. The score of 5 reflects partial circularity in the theoretical claims while acknowledging that the benchmark-driven performance demonstration is not circular.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method rests on two imported tools, InfoNCE's mutual information lower bound and SEM's CMI discrepancy weight, plus a new silhouette-based view quality weight. The free parameters are the hand-set loss weights, per-dataset network widths, and per-dataset training schedules. No new entities are introduced. The main unproven premise is that silhouette score reliably identifies the best anchoring view.

free parameters (5)
  • lambda (reconstruction loss weight) = 1.0
    Set to 1.0 for all datasets with no derivation; sensitivity analysis on two datasets shows robustness, but the value is still hand-chosen.
  • gamma (contrastive loss weight) = 1.0
    Set to 1.0 for all datasets; chosen by hand and tested only on Caltech6V7 and NUSWIDE.
  • per-dataset training schedule (pre-training epochs, CL iterations, CL epochs) = Varies per dataset; see Table 4
    Selected separately for each dataset, so results reflect tuned schedules rather than a single fixed protocol.
  • per-dataset encoder and MLP widths = See Table 3
    Network widths are manually specified per dataset; no ablation justifies these choices.
  • learning rate and batch size = 0.0003, 128
    Fixed across datasets by hand; standard choices but not derived.
assumptions (4)
  • standard math InfoNCE contrastive loss with a positive weight is a valid surrogate for mutual information maximization (Oord et al. 2018)
    Invoked in Theorem 1 to equate minimizing the dual-weighted loss with maximizing I(H_vi, H_vj); the equality step is a derivation gap.
  • domain assumption Silhouette coefficient on k-means cluster labels ranks view quality for contrastive anchoring
    Used in Eq. (8) to select the best view; no theoretical or empirical evidence shows that the highest-SI view is the optimal anchor.
  • domain assumption Cluster label distributions from each view's k-means encode shared semantic structure suitable for discrepancy weighting
    W_CMI in Eq. (13) is defined via CMI of view-specific cluster labels; this assumes the labels are meaningful and comparable.
  • domain assumption Representation degeneration can be mitigated by downweighting low-quality, high-discrepancy cross-views
    Motivates the dual weighting design; plausible but not proven independently.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DWCL: Dual-Weighted Contrastive Learning for Multi-View Clustering." pith.science (2026). https://pith.science/paper/AAXTV375

@misc{pith2026241117354,
  author       = {Pith},
  title        = {Pith review of: DWCL: Dual-Weighted Contrastive Learning for Multi-View Clustering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AAXTV375}},
  note         = {Machine review of arXiv:2411.17354}
}
read the original abstract

Multi-view contrastive clustering (MVCC) has gained significant attention for generating consistent clustering structures from multiple views through contrastive learning. However, most existing MVCC methods create cross-views by combining any two views, leading to a high volume of unreliable pairs. Furthermore, these approaches often overlook discrepancies in multi-view representations, resulting in representation degeneration. To address these challenges, we introduce a novel model called Dual-Weighted Contrastive Learning (DWCL) for Multi-View Clustering. Specifically, to reduce the impact of unreliable cross-views, we introduce an innovative Best-Other (B-O) contrastive mechanism that enhances the representation of individual views at a low computational cost. Furthermore, we develop a dual weighting strategy that combines a view quality weight, reflecting the quality of each view, with a view discrepancy weight. This approach effectively mitigates representation degeneration by downplaying cross-views that are both low in quality and high in discrepancy. We theoretically validate the efficiency of the B-O contrastive mechanism and the effectiveness of the dual weighting strategy. Extensive experiments demonstrate that DWCL outperforms previous methods across eight multi-view datasets, showcasing superior performance and robustness in MVCC. Specifically, our method achieves absolute accuracy improvements of 5.4\% and 5.6\% compared to state-of-the-art methods on the Caltech6V7 and MSRCv1 datasets, respectively.

Figures

Figures reproduced from arXiv: 2411.17354 by the authors.

Figure 1
Figure 1. Comparison of our DWCL with existing works in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The framework of DWCL. Silhouette coefficient (SI), the internal evaluation index of k-means clustering, is utilized to [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of DWCL with seven baseline methods [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: (a) Initial accuracy of each view on Caltech6V7. (b) Changes in weights [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: Results of ACC and NMI for Different Combinations [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Results of ACC and NMI for Different Combinations [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 5
Figure 5. Figure 5: Convergence Results of the Best View Compared to [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 8
Figure 8. Figure 8: Visual Comparison of Two Representative Multi-View Contrastive Learning Methods on the Fashion Dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Visualization of DWCL in the Contrastive Learning [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 7
Figure 7. Figure 7: Our analysis indicates that DWCL maintains robust cluster [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 48 canonical work pages

  1. [1]

    Unsu- pervised learning of visual features by contrasting cluster assignments

    Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A., 2020. Unsu- pervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems , 9912–9924

    work page 2020

  2. [2]

    Deep multiview clustering by contrasting cluster assignments, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Chen, J., Mao, H., Woo, W.L., Peng, X., 2023. Deep multiview clustering by contrasting cluster assignments, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 16752–16761

    work page 2023

  3. [3]

    A simple framework for contrastive learning of visual representations, in: International Conference on Machine Learning, pp

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G., 2020. A simple framework for contrastive learning of visual representations, in: International Conference on Machine Learning, pp. 1597–1607

    work page 2020

  4. [4]

    Support-vector networks

    Cortes, C., Vapnik, V ., 1995. Support-vector networks. Machine Learning , 273–297

    work page 1995

  5. [5]

    A novel ap- proach for effective multi-view clustering with information-theoretic perspective

    Cui, C., Ren, Y ., Pu, J., Li, J., Pu, X., Wu, T., Shi, Y ., He, L., 2024a. A novel ap- proach for effective multi-view clustering with information-theoretic perspective. Advances in Neural Information Processing Systems , 44847–44859

    work page

  6. [6]

    Dual contrast-driven deep multi-view clustering

    Cui, J., Li, Y ., Huang, H., Wen, J., 2024b. Dual contrast-driven deep multi-view clustering. IEEE Transactions on Image Processing , 4753–4764

    work page

  7. [7]

    Mvgan: Multi-view graph attention network for social event detection

    Cui, W., Du, J., Wang, D., Kou, F., Xue, Z., 2021. Mvgan: Multi-view graph attention network for social event detection. ACM Transactions on Intelligent Systems and Technology , 1–24

    work page 2021

  8. [8]

    Fei-Fei, L., Fergus, R., Perona, P., 2004. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pp. 178–178

    work page 2004

Show all 51 references
  1. [9]

    A bayesian hierarchical model for learning natural scene categories, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Fei-Fei, L., Perona, P., 2005. A bayesian hierarchical model for learning natural scene categories, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 524–531

    2005

  2. [10]

    Deep sparse rectifier neural networks, in: Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pp

    Glorot, X., Bordes, A., Bengio, Y ., 2011. Deep sparse rectifier neural networks, in: Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pp. 315–323

    2011

  3. [11]

    Mixgen: A new multi-modal data augmentation, in: Proceedings of the IEEE/CVF winter conference on applications of computer vision, pp

    Hao, X., Zhu, Y ., Appalaraju, S., Zhang, A., Zhang, W., Li, B., Li, M., 2023. Mixgen: A new multi-modal data augmentation, in: Proceedings of the IEEE/CVF winter conference on applications of computer vision, pp. 379–389

    2023

  4. [12]

    Algorithm as 136: A k-means clustering algorithm

    Hartigan, J.A., Wong, M.A., 1979. Algorithm as 136: A k-means clustering algorithm. Journal of the Royal Statistical Society. Series C (applied statistics) , 100–108

    1979

  5. [13]

    Momentum contrast for unsupervised visual representation learning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

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

    2020

  6. [14]

    Generalized information-theoretic multi-view clustering

    Huang, W., Yang, S., Cai, H., 2024. Generalized information-theoretic multi-view clustering. Advances in Neural Information Processing Systems , 58752–58764

    2024

  7. [15]

    Self-supervised graph attention networks for deep weighted multi-view clustering, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp

    Huang, Z., Ren, Y ., Pu, X., Huang, S., Xu, Z., He, L., 2023. Self-supervised graph attention networks for deep weighted multi-view clustering, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 7936–7943

    2023

  8. [16]

    Jiang, Y .G., Ye, G., Chang, S.F., Ellis, D., Loui, A.C., 2011. Consumer video understanding: A benchmark database and an evaluation of human and machine performance, in: Proceedings of the ACM International Conference on Multimedia Retrieval, pp. 1–8

    2011

  9. [17]

    Jin, J., Wang, S., Dong, Z., Liu, X., Zhu, E., 2023. Deep incomplete multi-view clustering with cross-view partial sample and prototype alignment, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11600–11609

    2023

  10. [18]

    Adam: A method for stochastic optimization

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

    2014 arXiv

  11. [19]

    Learning multiple layers of features from tiny images

    Krizhevsky, A., Hinton, G., et al., 2009. Learning multiple layers of features from tiny images. Handbook of Systemic Autoimmune Diseases , 1–16

    2009

  12. [20]

    Deep adversarial multi-view clustering network., in: International Joint Conference on Artificial Intelligence, p

    Li, Z., Wang, Q., Tao, Z., Gao, Q., Yang, Z., et al., 2019. Deep adversarial multi-view clustering network., in: International Joint Conference on Artificial Intelligence, p. 4

    2019

  13. [21]

    Dual contrastive prediction for incomplete multi-view representation learning

    Lin, Y ., Gou, Y ., Liu, X., Bai, J., Lv, J., Peng, X., 2022. Dual contrastive prediction for incomplete multi-view representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence , 4447–4461

    2022

  14. [22]

    Completer: Incomplete multi-view clustering via contrastive prediction, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Lin, Y ., Gou, Y ., Liu, Z., Li, B., Lv, J., Peng, X., 2021. Completer: Incomplete multi-view clustering via contrastive prediction, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11174–11183

    2021

  15. [23]

    Human action recognition and retrieval using sole depth information, in: Proceedings of the ACM International Conference on Multimedia, pp

    Lin, Y .C., Hu, M.C., Cheng, W.H., Hsieh, Y .H., Chen, H.M., 2012. Human action recognition and retrieval using sole depth information, in: Proceedings of the ACM International Conference on Multimedia, pp. 1053–1056

    2012

  16. [24]

    Cross-view approximation on grassmann manifold for multiview clustering

    Ma, Y ., Shen, X., Wu, D., Cao, J., Nie, F., 2024. Cross-view approximation on grassmann manifold for multiview clustering. IEEE Transactions on Neural Networks and Learning Systems , 1–6

    2024

  17. [25]

    Some methods for classification and analysis of multivariate observations, in: Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, pp

    MacQueen, J., et al., 1967. Some methods for classification and analysis of multivariate observations, in: Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, pp. 281–297

    1967

  18. [26]

    Representation learning with contrastive predictive coding

    Oord, A.v.d., Li, Y ., Vinyals, O., 2018. Representation learning with contrastive predictive coding. arXiv Preprint arXiv:1807.03748

    2018 arXiv

  19. [27]

    Multi-view contrastive graph clustering

    Pan, E., Kang, Z., 2021. Multi-view contrastive graph clustering. Advances in Neural Information Processing Systems , 2148–2159

    2021

  20. [28]

    Silhouettes: a graphical aid to the interpretation and validation of cluster analysis

    Rousseeuw, P.J., 1987. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics , 53–65

    1987

  21. [29]

    Scalable multi-view subspace clustering with unified anchors, in: Proceedings of the ACM International Conference on Multimedia, pp

    Sun, M., Zhang, P., Wang, S., Zhou, S., Tu, W., Liu, X., Zhu, E., Wang, C., 2021. Scalable multi-view subspace clustering with unified anchors, in: Proceedings of the ACM International Conference on Multimedia, pp. 3528–3536

    2021

  22. [30]

    Deep safe incomplete multi-view clustering: Theorem and algorithm, in: International Conference on Machine Learning, pp

    Tang, H., Liu, Y ., 2022a. Deep safe incomplete multi-view clustering: Theorem and algorithm, in: International Conference on Machine Learning, pp. 21090– 21110

  23. [31]

    Tang, H., Liu, Y ., 2022b. Deep safe multi-view clustering: Reducing the risk of clustering performance degradation caused by view increase, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 202–211

  24. [32]

    Trosten, D.J., Løkse, S., Jenssen, R., Kampffmeyer, M.C., 2023. On the ef- fects of self-supervision and contrastive alignment in deep multi-view clustering, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 23976–23985

    2023

  25. [33]

    One-step multi-view clustering with diverse representation

    Wan, X., Liu, J., Gan, X., Liu, X., Wang, S., Wen, Y ., Wan, T., Zhu, E., 2024a. One-step multi-view clustering with diverse representation. IEEE Transactions on Neural Networks and Learning Systems , 1–13

  26. [34]

    Wan, X., Liu, X., Liu, J., Wang, S., Wen, Y ., Liang, W., Zhu, E., Liu, Z., Zhou, L.,

  27. [35]

    Fast continual multi- view clustering with incomplete views

    Wan, X., Xiao, B., Liu, X., Liu, J., Liang, W., Zhu, E., 2024b. Fast continual multi- view clustering with incomplete views. IEEE Transactions on Image Processing , 2995–3008

  28. [36]

    Contrastive and attentive graph learning for multi-view clustering

    Wang, R., Li, L., Tao, X., Wang, P., Liu, P., 2022. Contrastive and attentive graph learning for multi-view clustering. Information Processing & Management , 102967

    2022

  29. [37]

    A self-representation method with local similarity preserving for fast multi-view outlier detection

    Wang, Y ., Chen, C., Lai, J., Fu, L., Zhou, Y ., Zheng, Z., 2023. A self-representation method with local similarity preserving for fast multi-view outlier detection. ACM Transactions on Knowledge Discovery from Data , 1–20

    2023

  30. [38]

    Scalable incomplete multi-view clustering with structure alignment, in: Proceedings of the ACM International Conference on Multimedia, pp

    Wen, Y ., Wang, S., Liang, K., Liang, W., Wan, X., Liu, X., Liu, S., Liu, J., Zhu, E., 2023. Scalable incomplete multi-view clustering with structure alignment, in: Proceedings of the ACM International Conference on Multimedia, pp. 3031–3040

    2023

  31. [39]

    Locus: Learning object classes with unsupervised segmentation, in: IEEE International Conference on Computer Vision, pp

    Winn, J., Jojic, N., 2005. Locus: Learning object classes with unsupervised segmentation, in: IEEE International Conference on Computer Vision, pp. 756– 763

    2005

  32. [40]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms

    Xiao, H., Rasul, K., V ollgraf, R., 2017. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv Preprint arXiv:1708.07747

    2017 arXiv

  33. [41]

    Self- weighted contrastive learning among multiple views for mitigating representation degeneration

    Xu, J., Chen, S., Ren, Y ., Shi, X., Shen, H., Niu, G., Zhu, X., 2023a. Self- weighted contrastive learning among multiple views for mitigating representation degeneration. Advances in Neural Information Processing Systems , 1119–1131

  34. [42]

    Adaptive feature projection with distribution alignment for deep incomplete multi-view clustering

    Xu, J., Li, C., Peng, L., Ren, Y ., Shi, X., Shen, H.T., Zhu, X., 2023b. Adaptive feature projection with distribution alignment for deep incomplete multi-view clustering. IEEE Transactions on Image Processing , 1354–1366

  35. [43]

    Xu, J., Ren, Y ., Tang, H., Pu, X., Zhu, X., Zeng, M., He, L., 2021. Multi-vae: Learning disentangled view-common and view-peculiar visual representations for multi-view clustering, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 9234–9243

    2021

  36. [44]

    Multi-level feature learning for contrastive multi-view clustering, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Xu, J., Tang, H., Ren, Y ., Peng, L., Zhu, X., He, L., 2022. 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

  37. [45]

    Anchor-sharing and clusterwise contrastive network for multiview representation learning

    Yan, W., Zhang, Y ., Tang, C., Zhou, W., Lin, W., 2024. Anchor-sharing and clusterwise contrastive network for multiview representation learning. IEEE Transactions on Neural Networks and Learning Systems , 1–11

    2024

  38. [46]

    Deep multiview collaborative clustering

    Yang, X., Deng, C., Dang, Z., Tao, D., 2021. Deep multiview collaborative clustering. IEEE Transactions on Neural Networks and Learning Systems , 516– 526

    2021

  39. [47]

    Dealmvc: Dual contrastive calibration for multi-view clustering, in: Proceedings of the ACM International Conference on Multimedia, pp

    Yang, X., Jiaqi, J., Wang, S., Liang, K., Liu, Y ., Wen, Y ., Liu, S., Zhou, S., Liu, X., Zhu, E., 2023. Dealmvc: Dual contrastive calibration for multi-view clustering, in: Proceedings of the ACM International Conference on Multimedia, pp. 337–346

    2023

  40. [48]

    Dvsai: Diverse view-shared anchors based incomplete multi-view clustering, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp

    Yu, S., Wang, S., Zhang, P., Wang, M., Wang, Z., Liu, Z., Fang, L., Zhu, E., Liu, X., 2024. Dvsai: Diverse view-shared anchors based incomplete multi-view clustering, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 16568–16577

    2024

  41. [49]

    A multiview clustering framework for detecting deceptive reviews

    Zhang, Y ., Wang, H., Stavrou, A., 2024. A multiview clustering framework for detecting deceptive reviews. Journal of Computer Security , 1–22

    2024

  42. [50]

    End-to-end adversarial-attention network for multi- modal clustering, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Zhou, R., Shen, Y .D., 2020. End-to-end adversarial-attention network for multi- modal clustering, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 14619–14628

    2020

  43. [2023]

    10078–10086

    Auto-weighted multi-view clustering for large-scale data, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 10078–10086

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.