Pith. sign in

REVIEW 4 major objections 4 minor 56 references

Towards Comprehensive Information-theoretic Multi-view Learning

T0 review · 4 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 This paper proposes a multi-view learning framework that keeps both shared and view-specific information and proves the combined representation stays predictively sufficient for the downstream label.

desk verdict The predictive-sufficiency theorem is assumed rather than proven, and the perfect MSRC-v1 score looks suspicious, but the common/unique IB+GK design is a sensible heuristic. read the letter →

arxiv 2509.02084 v1 pith:IDNYYKIV submitted 2025-09-02 cs.LG

Long Shi , Yunshan Ye , Wenjie Wang , Tao Lei , Yu Zhao , Gang Kou , Badong Chen This is my paper
classification cs.LG
keywords multi-viewlearninginformationbottleneckGács-Körnercommonpredictivesufficiencyview-uniquemutualrepresentationredundancy
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 proposes CIML, a multi-view learning framework that rejects the common 'multi-view redundancy' assumption that only information shared by all views matters for prediction. It learns two kinds of representation: a common representation extracted with Gács-Körner common information and then compressed by the information bottleneck, and per-view unique representations compressed separately while being pushed to be independent of the common part and of each other. The authors prove that the combined representation is predictively sufficient: if the original views fully determine the label, the learned joint representation retains essentially the same predictive information. On six benchmark datasets, CIML beats or matches state-of-the-art baselines, and ablation studies show both common and unique parts contribute.

What carries the argument

The load-bearing object is the joint representation Z=(Zc, Zu^(1),...,Zu^(v)), split into a common part Zc and per-view unique parts. Zc is learned by maximizing Gács-Körner common information, the largest latent variable reconstructable from every view, and then compressing through an information-bottleneck objective I(Zc;Y) - beta1 I(Zc;C). Each unique part Zu^(i) is learned with IB plus penalties minimizing I(Zu^(i);Zc) and cross-view I(Zu^(i);Zu^(j)); mutual-information terms are bounded variationally, with MINE used for the independence penalties. The argument's workhorse is the additivity identity I(Zc,Zu^1,...,Zu^v;Y) is approximately I(Zc;Y) plus the sum of I(Zu^i;Y), asserted to fol

What would settle it

Train CIML on a constructed multi-view dataset where the label is the XOR of one unique bit from view 1 and one unique bit from view 2, plus any shared bits. If the model makes the unique representations independent, a classifier reading Zc, Zu1, and Zu2 cannot recover the XOR correlation unless the joint representation preserves the pair, so accuracy will drop far below a classifier with access to raw views; measuring I(X;Y) versus I(Z;Y) directly would settle whether Eq. (B7) holds.

Watch

Extended reading notes

Core claim

The central claim is that a representation can use both the information shared by all views and the information unique to each view, and still retain all predictive information about the target. Concretely, Theorem 3.1 asserts that if the raw multi-view input X fully determines label Y, then the trained joint representation Z=(Zc,Zu^1,...,Zu^v) approximately determines Y, so H(Y|Z) is close to H(Y|X). The construction makes the claim work: Gács-Körner common information pins down the shared part, information bottleneck compresses both common and unique parts to task-relevant signal, and independence penalties separate the parts so their predictive contributions add.

Load-bearing premise

The proof assumes that once the common and per-view unique pieces are made mutually independent, their separate contributions to predicting the label sum exactly; independent pieces can still carry information about the label only when combined, so this additivity can fail.

Editorial extensions

If this is right

  • If predictive sufficiency holds, downstream classifiers can use the concatenated common-plus-unique representation instead of raw views without losing label information.
  • The multi-view redundancy assumption can be dropped without paying a predictive price; on reported benchmarks CIML meets or exceeds methods built on that assumption.
  • The framework is not limited to two views: Gács-Körner common information gives a shared representation across any number of views instead of pairwise averages.
  • Ablations show common and unique modules are complementary: removing either loss lowers accuracy, and on LandUse-21 the unique module contributes more than the common module.

Reading between the lines

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

  • If the additivity step in Eq. (B7) is tested directly, XOR-like label functions are the natural probe: two unique bits that are each useless alone but decisive together should reveal whether the independence penalties sacrifice joint predictive information.
  • The common/unique decomposition could be extended to incomplete multi-view learning, the paper's stated future direction, by using Gács-Körner common information as a reconstruction target for missing views; this follows from the modularity of the two loss terms.
  • A practitioner should verify that the MINE-based near-zero mutual-information penalties correspond to true near-independence at the dataset's finite sample size, since the predictive-sufficiency proof relies on those penalties actually holding.
  • The independence constraints create a representation where each view's unique contribution is individually decodable, which could be useful for interpretability and per-view attribution, though the paper does not explore this.
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 / 4 minor

Summary. The manuscript proposes CIML, a multi-view representation learning method that explicitly models both shared information across views and view-specific information. The common representation is learned by maximizing a heuristic Gács-Körner (GK) common-information objective followed by an information-bottleneck compression step; the unique representations are learned via IB objectives penalized by mutual information terms that encourage pairwise independence among unique representations and between unique and common representations. The central advertised result is Theorem 3.1, which claims that the concatenated joint representation Z=(Zc,Zu^1,...,Zu^v) is predictively sufficient for target Y. Experiments are reported on six datasets against a number of existing multi-view methods, together with ablation, hyperparameter, convergence, and visualization analyses. The main theoretical claim is, however, not established by the proof in Appendix B, and the GK interpretation of Eq. (1) is asserted rather than derived.

Significance. The motivation is timely: relaxing the multi-view redundancy assumption and treating unique information as predictive is an active research problem, and the proposed framework is modular enough to be of interest if its claims were sound. Strengths include the explicit common/unique decomposition, ablations, parameter sensitivity analysis, and release of code. However, the advertised predictive-sufficiency theorem is the paper’s main conceptual contribution, and the proof in Appendix B assumes the additivity that the theorem needs to show. The connection between Eq. (1) and Gács-Körner common information is also not justified. These are load-bearing issues rather than presentation defects. If the proof could be repaired under explicit nontrivial assumptions, the framework would be worth revisiting, but the manuscript as written does not provide a valid theoretical guarantee.

major comments (4)
  1. [Appendix B, Eqs. (B4)–(B5)] The proof’s load-bearing step is the assertion that, under the independence constraints, I(Zu^1,...,Zu^v; Y | Zc) ≈ Σ_i I(Zu^i; Y | Zc) ≈ Σ_i I(Zu^i; Y). This is false in general. Pairwise or even mutual independence among Zc and the Zu^i does not imply that their mutual information with Y decomposes additively, nor does it make the conditional mutual information equal to the marginal. A standard counterexample is Y = Zc XOR Zu^1 with Zc and Zu^1 independent Bernoulli(1/2): I(Zc;Y)=I(Zu^1;Y)=0 but I(Zc,Zu^1;Y)=1 bit, and I(Zu^1; Y | Zc)=1 bit. Thus Eq. (B5) is not a consequence of the training constraints.
  2. [Appendix B, Eq. (B7)] Even if Eq. (B5) were granted, Eq. (B7) states that I(Zc;Y) + Σ_i I(Zu^i;Y) ≈ I(X;Y). This is essentially the conclusion to be proven: it asserts that the joint representation retains approximately all predictive information from the original views. The training objectives in Eqs. (10)–(17) do not force this equality; the independence penalties at most enforce pairwise independence, and the IB terms are variational lower bounds. The theorem therefore reduces to an assumption about the learned representation rather than a guarantee, and its statement is close to tautological ('Z approximately determines Y if Z approximately determines Y').
  3. [Section III-B, Eq. (1)] The paper claims that maximizing H(C) − Σ_i ||fi(X^(i)) − C||^2 realizes Gács-Körner common information. This is not demonstrated. GK common information is a zero-error notion based on common randomness reconstructable from each source; an l2 reconstruction discrepancy does not by itself correspond to maximizing the entropy of a common part. No derivation, theorem, or reference in the manuscript establishes this equivalence. Since the common-representation module is a central component of the framework, the connection needs to be made precise or the claim should be weakened.
  4. [Section IV-A and IV-B] The experimental protocol is not sufficiently rigorous to support the claimed superiority. Hyperparameters β3 and β4 are tuned over predefined sets to 'achieve optimal performance' on the same datasets used for final reporting, with no validation split described and no significance tests. In such a setup, the reported gains (e.g., 100.00±0.00 on MSRC-v1) may reflect selection on test performance. A proper train/validation/test protocol and statistical comparison are needed before the empirical claims can be accepted.
minor comments (4)
  1. [Algorithm 1, Step 7] Step 7 says 'Calculate unique representation loss Lu by Eq. (18)', but the unique loss is in Eq. (17); Eq. (18) is the cross-entropy loss.
  2. [Throughout] Typographical errors include 'accouts', 'minmizing', 'uniqe', and 'down-steam'. 'NoisyMINST' in Table II and in the Section IV-B heading is inconsistent with 'NoisyMNIST' used elsewhere.
  3. [Equation (9)] The notation “Σ_{j=k}” in the second displayed line should presumably be “Σ_{k=1}”.
  4. [Figures 2–4] The figure captions are confusing: Fig. 4 appears to have duplicated subcaptions, and the caption lists multiple '(a)–(f)' blocks. Please clarify which panels refer to Zc and which to Zu.

Circularity Check

2 steps flagged · score 7.0 of 10

Predictive-sufficiency proof assumes its conclusion in Eq. (B7); Theorem 3.1 restates the definition.

  1. self definitional [Section III-F, Theorem 3.1 and Appendix B, Definition]
    "The joint representation Z = (Zc, Zu^(1), . . . ,Zu^(v)) is predictively sufficient for the target Y if, given that the original views X = (X^(1), · · ·, X^(v)) fully determine Y, the joint representation Z approximately determines Y."

    Appendix B defines predictive sufficiency as H(Y|Zc,Zu^1,...,Zu^v) ≈ H(Y|X^1,...,X^v). Under the theorem's premise that X fully determines Y, H(Y|X)=0, so 'Z approximately determines Y' is exactly H(Y|Z)≈0, i.e., the definition itself. The theorem therefore restates the definition as a conclusion rather than deriving it from the learning objective.

  2. self definitional [Appendix B, Eq. (B7)]
    "Therefore, under practical settings with small β1 and β2, the combined mutual information between Zc, Zu^(1), . . . ,Zu^(v) and Y closely matches that of the original views I(Zc; Y ) + Σv i=1 I(Z(i)u ; Y ) ≈ I(X^(1), . . . ,X^(v); Y )."

    Predictive sufficiency is equivalently I(Z;Y)≈I(X;Y). With Eq. (B5), Eq. (B7) states exactly I(Zc,Zu^1,...,Zu^v;Y)≈I(X^1,...,X^v;Y), which is the theorem's conclusion. This equality is not derived from the training losses in Eqs. (2), (11), or (12); the pairwise independence penalties only give I(Zu_i;Zc)≈0 and I(Zu_i;Zu_j)≈0, which do not imply the additivity used in B4/B5 or that the learned components jointly exhaust I(X;Y). Thus B7 is an assumed restatement of the target claim, not a proof of it.

full rationale

The experimental section is conventional and benchmarked against external methods, so that part is not circular. The circularity is confined to the paper's central theoretical claim of predictive sufficiency. Theorem 3.1 is already a restatement of the Appendix B definition: when X fully determines Y, 'Z approximately determines Y' is the defining condition H(Y|Z)≈H(Y|X). More importantly, the proof's load-bearing Eq. (B7) asserts that the mutual information of the learned joint representation with Y equals that of the original views. After the approximate decomposition in Eq. (B5), this is precisely the sufficiency equality the theorem is supposed to establish. That equality is not a consequence of the GK-IB objective or the independence penalties; it is assumed 'under practical settings.' The independence constraints enforce only pairwise zero mutual information and do not rule out synergistic label dependence, so the decomposition in B4/B5 is additionally unsupported. Consequently, the theoretical derivation reduces to an input assumption equivalent to its own conclusion. Score 7 reflects that the central proof is circular while the empirical benchmarking remains independent.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a decomposition of predictive information whose validity is assumed, on a naming of the common objective as Gács-Körner common information without derivation, and on hyperparameters tuned without a valid hold-out protocol.

free parameters (6)
  • beta_1 = 1e-4
    Trade-off in the common IB objective; fixed by hand to a small value.
  • beta_2 = 1e-4
    Trade-off in the unique IB objective; fixed by hand.
  • beta_3 = tuned over {10^0,...,10^3}
    Weight for common loss; tuned per dataset for best test accuracy without a specified validation split.
  • beta_4 = tuned over {10^-3,...,10^0}
    Weight for unique loss; tuned per dataset.
  • dim_Zc = not specified
    Dimensionality of common representation, selected via sensitivity analysis; no final value reported.
  • dim_Zu = not specified
    Dimensionality of each unique representation, selected via sensitivity analysis.
assumptions (5)
  • domain assumption The original views X fully determine the target Y.
    Stated in Theorem 3.1; in supervised classification this is often assumed, but it is not verified on the datasets.
  • ad hoc to paper Optimizing the independence penalties yields exact mutual independence between Zc and the Zu's and among the Zu's.
    Used in Appendix B to simplify I(Zu...;Y|Zc) into Σ I(Zu^i;Y); the paper provides no argument that the penalties are strong enough to enforce this.
  • ad hoc to paper The MI decomposition I(X;Y) ≈ I(Zc;Y) + Σ_i I(Zu^i;Y) holds.
    Eq. (B7); this is the central additivity assumption that is never derived, and it is false in general.
  • ad hoc to paper Eq. (1)'s objective max H(C) - Σ ||fi(Xi)-C||^2 realizes Gács-Körner common information.
    The paper names it GK common information but does not derive it from the GK definition.
  • domain assumption The variational bounds q and r are tight enough to make the approximated MI terms reliable for training.
    Standard practice in VIB-style methods; not verified here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Comprehensive Information-theoretic Multi-view Learning." pith.science (2026). https://pith.science/paper/IDNYYKIV

@misc{pith2026250902084,
  author       = {Pith},
  title        = {Pith review of: Towards Comprehensive Information-theoretic Multi-view Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IDNYYKIV}},
  note         = {Machine review of arXiv:2509.02084}
}
read the original abstract

Information theory has inspired numerous advancements in multi-view learning. Most multi-view methods incorporating information-theoretic principles rely an assumption called multi-view redundancy which states that common information between views is necessary and sufficient for down-stream tasks. This assumption emphasizes the importance of common information for prediction, but inherently ignores the potential of unique information in each view that could be predictive to the task. In this paper, we propose a comprehensive information-theoretic multi-view learning framework named CIML, which discards the assumption of multi-view redundancy. Specifically, CIML considers the potential predictive capabilities of both common and unique information based on information theory. First, the common representation learning maximizes Gacs-Korner common information to extract shared features and then compresses this information to learn task-relevant representations based on the Information Bottleneck (IB). For unique representation learning, IB is employed to achieve the most compressed unique representation for each view while simultaneously minimizing the mutual information between unique and common representations, as well as among different unique representations. Importantly, we theoretically prove that the learned joint representation is predictively sufficient for the downstream task. Extensive experimental results have demonstrated the superiority of our model over several state-of-art methods. The code is released on CIML.

Figures

Figures reproduced from arXiv: 2509.02084 by the authors.

Figure 1
Figure 1. Framework of CIML. III. PROPOSED METHODOLOGY In this section, we present the learning mechanisms of CIML, including both common and unique representations. Consider multi-view data {X(i)} v i=1 with n samples and m classes, where X(i) = [x (i) 1 , x (i) 2 , · · · , x (i) n ] ∈ R di×n repre￾sents the i-th view data with di dimension, with x (i) k denoting the k-th instance of the i-th view [PITH_FULL_IMAGE:figures/f… view at source ↗
Figure 2
Figure 2. Parameter sensitivity analysis with respect to [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Parameter sensitivity analysis with respect to [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Dimensionality sensitivity analysis with respect to [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Evolutionary curves of training loss and accuracy with respect to epoch. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: T-SNE visualization of our model on the LandUse-21, Caltech101-20, and Scene15 datasets. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 52 canonical work pages

  1. [1]

    Multi-view learning overview: Recent progress and new challenges,

    J. Zhao, X. Xie, X. Xu, and S. Sun, “Multi-view learning overview: Recent progress and new challenges,” Information Fusion, vol. 38, pp. 43–54, 2017

    work page 2017

  2. [2]

    Trusted multi-view classi- fication with dynamic evidential fusion,

    Z. Han, C. Zhang, H. Fu, and J. T. Zhou, “Trusted multi-view classi- fication with dynamic evidential fusion,” IEEE transactions on pattern analysis and machine intelligence , vol. 45, no. 2, pp. 2551–2566, 2022

    work page 2022

  3. [3]

    Reliable conflictive multi-view learning,

    C. Xu, J. Si, Z. Guan, W. Zhao, Y . Wu, and X. Gao, “Reliable conflictive multi-view learning,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 14, 2024, pp. 16 129–16 137

    work page 2024

  4. [4]

    Generalized Trusted Multi-view Classification Framework with Hierarchical Opinion Aggregation

    L. Shi, C. Tang, H. Deng, C. Xu, L. Xing, and B. Chen, “Generalized trusted multi-view classification framework with hierarchical opinion aggregation,” arXiv preprint arXiv:2411.03713 , 2024

    work page Pith review arXiv 2024

  5. [5]

    Discovering common in- formation in multi-view data,

    Q. Zhang, M. Lu, S. Yu, J. Xin, and B. Chen, “Discovering common in- formation in multi-view data,” Information Fusion, vol. 108, p. 102400, 2024

    work page 2024

  6. [6]

    Late fusion incomplete multi-view clustering,

    X. Liu, X. Zhu, M. Li, L. Wang, C. Tang, J. Yin, D. Shen, H. Wang, and W. Gao, “Late fusion incomplete multi-view clustering,” IEEE transactions on pattern analysis and machine intelligence , vol. 41, no. 10, pp. 2410–2423, 2018

    work page 2018

  7. [7]

    Dual contrastive prediction for incomplete multi-view representation learning,

    Y . Lin, Y . Gou, X. Liu, J. Bai, J. Lv, and X. Peng, “Dual contrastive prediction for incomplete multi-view representation learning,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 4, pp. 4447–4461, 2022

    work page 2022

  8. [8]

    Multi-view attributed graph clustering,

    Z. Lin, Z. Kang, L. Zhang, and L. Tian, “Multi-view attributed graph clustering,” IEEE Transactions on knowledge and data engineering , vol. 35, no. 2, pp. 1872–1880, 2021

    work page 2021

Show all 56 references
  1. [9]

    Deep variational information bottleneck,

    A. A. Alemi, I. Fischer, J. V . Dillon, and K. Murphy, “Deep variational information bottleneck,” arXiv preprint arXiv:1612.00410 , 2016

    2016 arXiv

  2. [10]

    Information-theoretic methods in deep neural networks: Recent advances and emerging opportunities

    S. Yu, L. G. S. Giraldo, and J. C. Pr ´ıncipe, “Information-theoretic methods in deep neural networks: Recent advances and emerging opportunities.” in IJCAI, 2021, pp. 4669–4678

    2021

  3. [11]

    A survey on information bottleneck,

    S. Hu, Z. Lou, X. Yan, and Y . Ye, “A survey on information bottleneck,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2024

    2024

  4. [12]

    Learning robust representations via multi-view information bottleneck,

    M. Federici, A. Dutta, P. Forr ´e, N. Kushman, and Z. Akata, “Learning robust representations via multi-view information bottleneck,” 2020

    2020

  5. [13]

    Multi-view information- bottleneck representation learning,

    Z. Wan, C. Zhang, P. Zhu, and Q. Hu, “Multi-view information- bottleneck representation learning,” in Proceedings of the AAAI confer- ence on artificial intelligence , vol. 35, no. 11, 2021, pp. 10 085–10 092

    2021

  6. [14]

    Dmib: Dual-correlated multivariate in- formation bottleneck for multiview clustering,

    S. Hu, Z. Shi, and Y . Ye, “Dmib: Dual-correlated multivariate in- formation bottleneck for multiview clustering,” IEEE Transactions on Cybernetics, vol. 52, no. 6, pp. 4260–4274, 2020

    2020

  7. [15]

    Self-supervised information bottleneck for deep multi-view subspace clustering,

    S. Wang, C. Li, Y . Li, Y . Yuan, and G. Wang, “Self-supervised information bottleneck for deep multi-view subspace clustering,” IEEE Transactions on Image Processing , vol. 32, pp. 1555–1567, 2023

    2023

  8. [16]

    A novel approach for effective multi-view clustering with information- theoretic perspective,

    C. Cui, Y . Ren, J. Pu, J. Li, X. Pu, T. Wu, Y . Shi, and L. He, “A novel approach for effective multi-view clustering with information- theoretic perspective,” Advances in Neural Information Processing Sys- tems, vol. 36, 2024

    2024

  9. [17]

    Multi-view subspace clustering,

    H. Gao, F. Nie, X. Li, and H. Huang, “Multi-view subspace clustering,” in Proceedings of the IEEE international conference on computer vision, 2015, pp. 4238–4246

    2015

  10. [18]

    Robust subspace clustering by logarithmic hyperbolic cosine function,

    L. Cao, L. Shi, J. Wang, Z. Yang, and B. Chen, “Robust subspace clustering by logarithmic hyperbolic cosine function,” IEEE Signal Processing Letters, vol. 30, pp. 508–512, 2023

    2023

  11. [19]

    Generalized latent multi-view subspace clustering,

    C. Zhang, H. Fu, Q. Hu, X. Cao, Y . Xie, D. Tao, and D. Xu, “Generalized latent multi-view subspace clustering,” IEEE transactions on pattern analysis and machine intelligence , vol. 42, no. 1, pp. 86–99, 2018

    2018

  12. [20]

    Multi-view cluster- ing in latent embedding space,

    M.-S. Chen, L. Huang, C.-D. Wang, and D. Huang, “Multi-view cluster- ing in latent embedding space,” in Proceedings of the AAAI conference on artificial intelligence , vol. 34, no. 04, 2020, pp. 3513–3520

    2020

  13. [21]

    Enhanced latent multi-view subspace clustering,

    L. Shi, L. Cao, J. Wang, and B. Chen, “Enhanced latent multi-view subspace clustering,” IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  14. [22]

    On unify- ing multi-view self-representations for clustering by tensor multi-rank minimization,

    Y . Xie, D. Tao, W. Zhang, Y . Liu, L. Zhang, and Y . Qu, “On unify- ing multi-view self-representations for clustering by tensor multi-rank minimization,” International Journal of Computer Vision , vol. 126, pp. 1157–1179, 2018

    2018

  15. [23]

    Unified tensor framework for incomplete multi-view clustering and missing-view inferring,

    J. Wen, Z. Zhang, Z. Zhang, L. Zhu, L. Fei, B. Zhang, and Y . Xu, “Unified tensor framework for incomplete multi-view clustering and missing-view inferring,” in Proceedings of the AAAI conference on artificial intelligence, vol. 35, no. 11, 2021, pp. 10 273–10 281

    2021

  16. [24]

    Tensorized and compressed multi-view subspace clustering via structured constraint,

    W. Chang, H. Chen, F. Nie, R. Wang, and X. Li, “Tensorized and compressed multi-view subspace clustering via structured constraint,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2024

    2024

  17. [25]

    Flexible tensor learning for multi- view clustering with markov chain,

    Y . Qin, Z. Tang, H. Wu, and G. Feng, “Flexible tensor learning for multi- view clustering with markov chain,” IEEE Transactions on Knowledge and Data Engineering , vol. 36, no. 4, pp. 1552–1565, 2023

    2023

  18. [26]

    Logarithmic schatten- p p norm minimization for tensorial multi-view subspace clustering,

    J. Guo, Y . Sun, J. Gao, Y . Hu, and B. Yin, “Logarithmic schatten- p p norm minimization for tensorial multi-view subspace clustering,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 3, pp. 3396–3410, 2022

    2022

  19. [27]

    Tensorized tri-factor decomposition for multi-view clustering,

    R. Wang, Q. Gao, M. Yang, and Q. Wang, “Tensorized tri-factor decomposition for multi-view clustering,” IEEE Transactions on Circuits and Systems for Video Technology , 2025

    2025

  20. [28]

    Gmc: Graph-based multi-view cluster- ing,

    H. Wang, Y . Yang, and B. Liu, “Gmc: Graph-based multi-view cluster- ing,” IEEE Transactions on Knowledge and Data Engineering , vol. 32, no. 6, pp. 1116–1129, 2019

    2019

  21. [29]

    Sample-level multi- view graph clustering,

    Y . Tan, Y . Liu, S. Huang, W. Feng, and J. Lv, “Sample-level multi- view graph clustering,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 23 966–23 975

    2023

  22. [30]

    Multi-view contrastive graph clustering,

    E. Pan and Z. Kang, “Multi-view contrastive graph clustering,” Advances in neural information processing systems, vol. 34, pp. 2148–2159, 2021

    2021

  23. [31]

    On deep multi- view representation learning,

    W. Wang, R. Arora, K. Livescu, and J. Bilmes, “On deep multi- view representation learning,” in International conference on machine learning. PMLR, 2015, pp. 1083–1092

    2015

  24. [32]

    Completer: Incomplete multi-view clustering via contrastive prediction,

    Y . Lin, Y . Gou, Z. Liu, B. Li, J. Lv, and X. Peng, “Completer: Incomplete multi-view clustering via contrastive prediction,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 11 174–11 183

    2021

  25. [33]

    Nim-nets: Noise-aware incomplete multi-view learning networks,

    Y . Qin, C. Qin, X. Zhang, D. Qi, and G. Feng, “Nim-nets: Noise-aware incomplete multi-view learning networks,” IEEE Transactions on Image Processing, vol. 32, pp. 175–189, 2022

    2022

  26. [34]

    Spectral embedding fusion for incomplete multiview clustering,

    J. Chen, Y . Chen, Z. Wang, H. Zhang, and X. Peng, “Spectral embedding fusion for incomplete multiview clustering,” IEEE Transactions on Image Processing, vol. 33, pp. 4116–4130, 2024

    2024

  27. [35]

    Tensor-based unsupervised feature selection for error-robust handling of unbalanced incomplete multi-view data,

    X. Yang, H. Che, and M.-F. Leung, “Tensor-based unsupervised feature selection for error-robust handling of unbalanced incomplete multi-view data,” Information Fusion, vol. 114, p. 102693, 2025

    2025

  28. [36]

    Anchor graph network for incomplete multiview clustering,

    Y . Fu, Y . Li, Q. Huang, J. Cui, and J. Wen, “Anchor graph network for incomplete multiview clustering,” IEEE Transactions on Neural Networks and Learning Systems , 2024

    2024

  29. [37]

    Reliable representation learning for incomplete multi-view missing multi-label classification,

    C. Liu, J. Wen, Y . Xu, B. Zhang, L. Nie, and M. Zhang, “Reliable representation learning for incomplete multi-view missing multi-label classification,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  30. [38]

    How does information bottleneck help deep learning?

    K. Kawaguchi, Z. Deng, X. Ji, and J. Huang, “How does information bottleneck help deep learning?” in International Conference on Machine Learning. PMLR, 2023, pp. 16 049–16 096

    2023

  31. [39]

    Generalized information-theoretic multi-view clustering,

    W. Huang, S. Yang, and H. Cai, “Generalized information-theoretic multi-view clustering,” Advances in neural information processing sys- tems, vol. 36, pp. 58 752–58 764, 2023

    2023

  32. [40]

    Differentiable information bot- tleneck for deterministic multi-view clustering,

    X. Yan, Z. Jin, F. Han, and Y . Ye, “Differentiable information bot- tleneck for deterministic multi-view clustering,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 27 435–27 444

    2024

  33. [41]

    Multiview representation learning via information-theoretic optimization,

    W. Yan, S. Yao, C. Tang, and W. Zhou, “Multiview representation learning via information-theoretic optimization,” IEEE Transactions on Neural Networks and Learning Systems , 2025

    2025

  34. [42]

    Self- supervised weighted information bottleneck for multi-view clustering,

    Z. Lou, C. Zhang, H. Xue, Y . Ye, Q. Zhou, and S. Hu, “Self- supervised weighted information bottleneck for multi-view clustering,” in Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, 2024, pp. 4643–4650

    2024

  35. [43]

    Zero-error information and applications in cryptography,

    S. Wolf and J. Wultschleger, “Zero-error information and applications in cryptography,” in Information Theory Workshop . IEEE, 2004, pp. 1–6

    2004

  36. [44]

    Common information is far less than mutual information

    P. G ´acs, J. Korner et al., “Common information is far less than mutual information.” Problems of Control and Information Theory , vol. 2, pp. 149–162, 1973

    1973

  37. [45]

    Gacs-korner common information variational autoencoder,

    M. Kleinman, A. Achille, S. Soatto, and J. Kao, “Gacs-korner common information variational autoencoder,” Advances in Neural Information Processing Systems, vol. 36, pp. 66 020–66 043, 2023

    2023

  38. [46]

    Learning representations by maximizing mutual information across views,

    P. Bachman, R. D. Hjelm, and W. Buchwalter, “Learning representations by maximizing mutual information across views,” Advances in neural information processing systems , vol. 32, 2019

    2019

  39. [47]

    Variational recurrent auto- encoders,

    O. Fabius and J. R. Van Amersfoort, “Variational recurrent auto- encoders,” arXiv preprint arXiv:1412.6581 , 2014

    2014 arXiv

  40. [48]

    Mine: mutual information neural estimation,

    M. I. Belghazi, A. Baratin, S. Rajeswar, S. Ozair, Y . Bengio, A. Courville, and R. D. Hjelm, “Mine: mutual information neural estimation,” arXiv preprint arXiv:1801.04062 , 2018. JOURNAL OF LATEX CLASS FILES 12

    2018 arXiv

  41. [49]

    Factorized contrastive learning: Going beyond multi- view redundancy,

    P. P. Liang, Z. Deng, M. Q. Ma, J. Y . Zou, L.-P. Morency, and R. Salakhutdinov, “Factorized contrastive learning: Going beyond multi- view redundancy,” Advances in Neural Information Processing Systems , vol. 36, pp. 32 971–32 998, 2023

    2023

  42. [50]

    Nus-wide: a real-world web image database from national university of singapore,

    T.-S. Chua, J. Tang, R. Hong, H. Li, Z. Luo, and Y . Zheng, “Nus-wide: a real-world web image database from national university of singapore,” in Proceedings of the ACM international conference on image and video retrieval, 2009, pp. 1–9

    2009

  43. [51]

    Deep canonical correlation analysis,

    G. Andrew, R. Arora, J. Bilmes, and K. Livescu, “Deep canonical correlation analysis,” in International conference on machine learning . PMLR, 2013, pp. 1247–1255

    2013

  44. [52]

    Uncertainty-aware multi-view representation learning,

    Y . Geng, Z. Han, C. Zhang, and Q. Hu, “Uncertainty-aware multi-view representation learning,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 9, 2021, pp. 7545–7553

    2021

  45. [53]

    Interpretable graph convolutional network for multi-view semi-supervised learning,

    Z. Wu, X. Lin, Z. Lin, Z. Chen, Y . Bai, and S. Wang, “Interpretable graph convolutional network for multi-view semi-supervised learning,” IEEE Transactions on Multimedia , vol. 25, pp. 8593–8606, 2023

    2023

  46. [54]

    Multi- view fuzzy classification with subspace clustering and information granules,

    X. Hu, X. Liu, W. Pedrycz, Q. Liao, Y . Shen, Y . Li, and S. Wang, “Multi- view fuzzy classification with subspace clustering and information granules,” IEEE Transactions on Knowledge and Data Engineering , vol. 35, no. 11, pp. 11 642–11 655, 2022

    2022

  47. [55]

    Deep incomplete multi-view learning network with insufficient label information,

    Z. Jiang, T. Luo, and X. Liang, “Deep incomplete multi-view learning network with insufficient label information,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 11, 2024, pp. 12 919– 12 927

    2024

  48. [56]

    Visualizing data using t-sne

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.” Journal of machine learning research , vol. 9, no. 11, 2008

    2008

Pith tools

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