Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Multi-View Factorizing and Disentangling: A Novel Framework for Incomplete Multi-View Multi-Label Classification

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Two-stage factorization outperforms ten baselines on incomplete multi-view multi-label classification.

desk verdict A solid empirical iMvMLC framework with a real but unaddressed train/test shift in its two-stage design; worth reviewing, needs a reproducibility pass. read the letter →

arxiv 2501.06524 v2 pith:HWHNCWTX submitted 2025-01-11 cs.CV

Wulin Xie , Lian Zhao , Jiang Long , Xiaohuan Lu , Bingyan Nie This is my paper · ORCID
classification cs.CV
keywords incompletemulti-viewmulti-labelclassificationview-consistentandview-specificfactorizationmaskedcross-viewpredictiongraphdisentanglinglosstwo-stagelearningsemanticcontrastiverepresentationdisentanglement
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes MVFD, a two-stage framework for incomplete multi-view multi-label classification (iMvMLC). Its central claim is that multi-view representations should not be learned jointly in an end-to-end model; instead, view-consistent and view-specific factors should be learned separately, first under random masking and then with a graph disentangling loss. The authors argue that this factorization avoids the sub-optimal min-max trade-off and redundancy of coupled end-to-end learning, and they report that MVFD outperforms ten comparison methods on all six metrics across five datasets at a 50% missing-view and missing-label rate. A sympathetic reader would care because incomplete views and labels are common in practice, and the proposed losses are designed to operate under arbitrary missing patterns.

What carries the argument

The load-bearing mechanism is the factorized two-stage paradigm: stage one learns view-consistent representations under a masked cross-view prediction strategy with three sub-objectives (shared-information extraction via reconstruction, intra-view redundancy elimination via semantic contrastive loss, and task-relevant information preservation via cross-entropy), and stage two freezes those encoders and learns view-specific representations through an autoencoder guided by the graph disentangling loss. The graph disentangling loss, Eq. (10), is the central identity: it sums, per sample, the cosine similarity between the consistent representation and each view-specific feature, plus the cosine similarity between view-specific features of different views, masked by view-availability indicators. Minimizing it is what the paper claims fully eliminates redundancy between the two factor sets. The final classification fuses the factors with the star operation $Z = \mathrm{Sigmoid}(\bar{S}) \odot \bar{C}$.

What would settle it

Measure the stage-one reconstruction loss $\ell_{cp}$ on unmasked inputs immediately after stage-one training: if the frozen consistent encoders have overfit to the masking pattern, the unmasked reconstruction error will be substantially higher than the masked error, and the graph disentangling loss in stage two operates on distorted representations. A second check is to run MVFD with the mask ratio $\delta=0$; if performance drops sharply, the masking strategy, not the factorization, is doing the work.

Watch

Extended reading notes

Core claim

MVFD factorizes each view's representation into a view-consistent factor $\hat{C}$ and a view-specific factor $\bar{S}$, learned in two stages. In stage one, randomly masked inputs $\bar{X}^{(v)}$ pass through consistent encoders $E_c^{(v)}$, are averaged by availability into $\hat{C}$, and are trained with a masked cross-view prediction loss $\ell_{cp}$, a semantic contrastive loss $\ell_{sc}$ on per-view predictions, and a task-relevant cross-entropy loss $\ell_{ce1}$. In stage two, the consistent encoders are frozen, a separate view-specific autoencoder extracts $S^{(v)}$, and a graph disentangling loss $\ell_{gd}$ minimizes cosine similarity between $\bar{C}$ and each $S^{(v)}$ and between different views' specific features. The final prediction uses the star operation $Z = \mathrm{Sigmoid}(\bar{S}) \odot \bar{C}$. The paper claims this design beats all ten baselines on all metrics on Corel5k, Pascal07, ESPGame, IAPRTC12, and MIRFLICKR at 50% missingness, and that ablation shows every loss component and the two-stage paradigm contribute.

Load-bearing premise

The consistent encoders are trained on randomly masked inputs in stage one, then frozen and run on unmasked inputs in stage two, and the method assumes these representations remain valid under that input-distribution shift.

Editorial extensions

If this is right

  • If the two-stage claim holds, end-to-end joint optimization is not necessary for iMvMLC; factorizing the representation first and disentangling second is a viable alternative that avoids the described min-max sub-optimality.
  • The graph disentangling loss is proposed as an adaptable component that can be dropped into existing multi-view representation learning methods to reduce redundancy between shared and specific factors.
  • The framework is designed to handle arbitrary missing-view and missing-label patterns, so it generalizes beyond fixed incomplete ratios.
  • Ablation results indicate each loss term contributes and the two-stage paradigm outperforms the one-stage consolidated version, so the decomposition itself, not any single loss, carries the performance.
  • On fully observed data, the method remains competitive, suggesting the factorization does not depend on missingness to help.

Reading between the lines

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

  • Because the graph disentangling loss only penalizes pairwise cosine similarity, 'fully eliminates redundancy' is stronger than what the loss enforces; a stricter independence test (e.g., Hilbert-Schmidt independence criterion on the learned factors) would be needed to verify complete disentanglement.
  • The two-stage design freezes encoders across a masked/unmasked distribution shift; if that shift is harmful, fine-tuning or adapting the consistent encoders during stage two would be a natural extension the paper does not explore.
  • All six views are handcrafted features; whether the factorization and disentangling losses transfer to views produced by deep neural encoders is untested and would be a direct next experiment.
  • The star operation's contribution is not isolated; replacing it with concatenation or summation in the same framework would test whether the fusion rule or the factorization drives the gains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes MVFD, a two-stage framework for incomplete multi-view multi-label classification (iMvMLC). Stage 1 learns view-consistent representations through masked cross-view prediction, a semantic contrastive loss on per-view predictions, and a task-relevant cross-entropy loss. Stage 2 freezes the stage-1 consistent encoders, trains view-specific encoders/decoders with a reconstruction loss, and applies a graph disentangling loss that penalizes cosine similarity between consistent and view-specific features and between view-specific features from different views. The final classification feature is the element-wise product of the sigmoid of the view-specific representation and the consistent representation. Experiments on five public datasets with 50% missing views and labels compare against ten methods and report gains across metrics, together with ablations and hyperparameter sensitivity figures.

Significance. The two-stage factorization is an interesting and timely design choice that avoids the joint min-max optimization of many end-to-end models. The loss design is coherent, and the experimental evaluation is broad: it uses five external public datasets, ten comparison methods, standard deviations, missing-rate analyses, and ablations, so the central claim is not circular. If the claims hold, MVFD would be a competitive baseline for iMvMLC. However, the claim that the graph disentangling loss 'fully eliminates' redundancy is stronger than what a cosine-similarity penalty can establish, and the paper omits several settings needed to reproduce the main numbers.

major comments (3)
  1. [Section 3.3, Eq. (9)] The two-stage design assumes that the consistent encoders E_c(v), trained exclusively on masked inputs via Eq. (1), produce valid representations when applied to the original unmasked X(v) in Eq. (9). The paper does not discuss or evaluate this train/test distribution shift. If E_c(v) has learned to rely on the zero-block mask pattern, then the \bar{C} used in the graph disentangling loss and in Eq. (12) is not the representation optimized by the masked consistent prediction loss in Eq. (3). Figure 6 only shows that the similarity loss decreases during stage 2; it does not demonstrate that the stage-2 \bar{C} is aligned with the stage-1 consistent representation. I would like to see an explicit experiment or discussion, for example by measuring the agreement between C'(v) computed from masked and unmasked inputs, or by briefly adapting the frozen encoder to unmasked inputs before computing \bar{C}.
  2. [Section 4.3] The values of the loss coefficients α, β, γ, λ and the mask ratio δ are not reported. Fig. 5 gives only approximate optimal ranges for α, β, γ, λ, and Eq. (1) defines δ without ever stating its value. Since all four coefficients appear in the final losses (Eqs. (7) and (14)), the main results in Table 1 and the ablation in Table 2 are not reproducible without these settings. Please provide a table with the exact configuration used for each dataset, including the mask ratio and the network architecture details such as encoder depth and embedding dimension.
  3. [Section 3.3, Eq. (10)] The graph disentangling loss minimizes cosine similarities between \bar{C} and S(v) and between S(v) and S(u). Minimizing cosine similarity does not imply that redundant information is 'fully eliminated': two vectors can have zero cosine similarity while still sharing substantial information through nonlinear dependencies or higher-order statistics. The abstract and the contributions state that the loss 'fully reduces' or 'fully eliminates' redundancy, which is not supported by the loss definition or by Fig. 6, since Fig. 6 only shows a decrease in the very similarity term being minimized. I suggest softening the claim or adding a quantitative redundancy measure, such as an estimate of mutual information or a normalized cross-covariance, evaluated on the learned representations.
minor comments (6)
  1. [Section 3.2, Eq. (7)] The sentence before Eq. (7) says 'all loss functions in the second stage becomes', but Eq. (7) defines ℓ_stage1; this should be corrected to refer to the first stage.
  2. [Section 4.2, first bullet] The claim that MVFD 'outperforms comparison methods on all metrics' is not strictly supported by Table 1, which shows exact ties on 1-HL for several datasets (for example, Corel5k has 0.988 for both DICNet and MVFD). Please qualify the claim.
  3. [Fig. 5 caption] The caption states that panels (c) and (d) show AP values for α and β, but the text describes them as showing γ and λ; please correct the mismatch.
  4. [Section 3.2, Eq. (4)] The sentence about positive and negative instances is confused: the anchor is indexed by view v, so the number of positive views is m−1, not N−1, and the description of the negative set is unclear. Please revise the sentence to match the formula.
  5. [Section 3.2, Proposition 1] Proposition 1 is used as a motivation but is not proved; it would be more accurate to present it as a definition or assumption rather than as a formal proposition.
  6. [Section 4.1] The text says that 'for all metrics, higher values indicate better performance', but Ranking Loss, OneError, and Coverage are originally lower-is-better metrics; the table reports their complements (1-RL, 1-OE, 1-Cov). This should be stated explicitly to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical SOTA claim is supported by external benchmarks and comparison methods, and the design components are defined operationally rather than by the results they are said to predict.

full rationale

The paper's central claim is empirical: MVFD outperforms ten comparison methods on five public datasets (Table 1) using standard metrics. This is tested against independent published methods and does not reduce to the paper's own inputs. The consistent representation is defined by a reconstruction criterion and learned through explicit losses: masked consistent prediction (Eq. 3), semantic contrastive loss (Eq. 4), and cross-entropy (Eq. 6); the view-specific representation is learned through a separate autoencoder reconstruction (Eq. 8) and a cosine-similarity graph disentangling loss (Eq. 10). No fitted parameter is renamed as a prediction; the only tuned quantities are the loss weights, and the ablation study (Table 2) removes each component to check its contribution. The two-stage training does contain a train/test distribution shift: stage-1 encoders are optimized on masked inputs and applied to unmasked inputs in Eq. (9). That is a correctness risk, not a circularity, because it does not make the result equivalent to an input by construction. Figure 6 shows that the graph disentangling loss decreases the cosine similarities it explicitly minimizes; this is a sanity check that the optimizer is doing what the loss defines, not an independent derived prediction, and it is not load-bearing for the benchmark comparison. Self-citations ([22], [23], [26], [38]) appear only in related-work enumeration and are not used to justify the method's core assertions. No circular step meeting the quoted-equation standard is present.

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

The framework uses standard representation-learning constructs (consistent and specific factors); no new physical or mathematical entities are postulated.

free parameters (5)
  • alpha (α) = not reported; optimal range [0.5,1] on Corel5k and [0.05,0.1] on Pascal07 (Fig. 5)
    Weight for masked consistent prediction loss ℓcp in Eq. (7); chosen by grid search, affects performance.
  • beta (β) = not reported; optimal range [0.05,0.5] on Corel5k and [0.05,0.1] on Pascal07
    Weight for semantic contrastive loss ℓsc; tuned per dataset.
  • gamma (γ) = not reported; optimal range [0.1,0.5] on Corel5k and [0.05,1] on Pascal07
    Weight for view-specific reconstruction loss ℓrec in Eq. (14).
  • lambda (λ) = not reported; optimal range [0.005,0.01] on Corel5k and [0.05,0.5] on Pascal07
    Weight for graph disentangling loss ℓgd.
  • mask ratio delta (δ) = not reported or studied
    Determines mask length l(v) = d_v * δ in Eq. (1); critical to the masked prediction objective, but no sensitivity analysis is given.
assumptions (4)
  • domain assumption Each observation x(v) can be reconstructed from a mapping f(v)(c) (Proposition 1).
    Invoked in Section 3.2 to justify the masked reconstruction loss; assumes a shared latent c suffices to generate each view.
  • domain assumption Views are conditionally independent given the shared multi-view representation (from [41] and [11]).
    Used in Section 3.2 to define the consistent representation and to justify cross-view reconstruction; not verified on the datasets.
  • standard math Minimizing the cross-entropy loss is equivalent to minimizing -I(Ĉ;Y).
    Used in Section 3.2 'Task-relevant information preserving'; a standard equivalence cited from [13], not derived in the paper.
  • domain assumption Missing views and labels are missing completely at random (MCAR).
    The experimental protocol creates random missingness with probabilities 0.5; the method does not model non-random missingness, which is common in practice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-View Factorizing and Disentangling: A Novel Framework for Incomplete Multi-View Multi-Label Classification." pith.science (2026). https://pith.science/paper/HWHNCWTX

@misc{pith2026250106524,
  author       = {Pith},
  title        = {Pith review of: Multi-View Factorizing and Disentangling: A Novel Framework for Incomplete Multi-View Multi-Label Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HWHNCWTX}},
  note         = {Machine review of arXiv:2501.06524}
}
read the original abstract

Multi-view multi-label classification (MvMLC) has recently garnered significant research attention due to its wide range of real-world applications. However, incompleteness in views and labels is a common challenge, often resulting from data collection oversights and uncertainties in manual annotation. Furthermore, the task of learning robust multi-view representations that are both view-consistent and view-specific from diverse views still a challenge problem in MvMLC. To address these issues, we propose a novel framework for incomplete multi-view multi-label classification (iMvMLC). Our method factorizes multi-view representations into two independent sets of factors: view-consistent and view-specific, and we correspondingly design a graph disentangling loss to fully reduce redundancy between these representations. Additionally, our framework innovatively decomposes consistent representation learning into three key sub-objectives: (i) how to extract view-shared information across different views, (ii) how to eliminate intra-view redundancy in consistent representations, and (iii) how to preserve task-relevant information. To this end, we design a robust task-relevant consistency learning module that collaboratively learns high-quality consistent representations, leveraging a masked cross-view prediction (MCP) strategy and information theory. Notably, all modules in our framework are developed to function effectively under conditions of incomplete views and labels, making our method adaptable to various multi-view and multi-label datasets. Extensive experiments on five datasets demonstrate that our method outperforms other leading approaches.

Figures

Figures reproduced from arXiv: 2501.06524 by the authors.

Figure 1
Figure 1. An overview of our MVFD framework. In the first age, we first randomly masked fragments of input features. Then we factorize [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. We define a complete multi-view data contain view [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Experimental results of eleven methods on three datasets without any missing views or labels. The worst results are indicated at [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Performance about the comparison with SOTA methods [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The AP values for hyper-parameters α and β on the Corel5k (Fig5. a) and Pascal07 (Fig5. b) datasets; AP values for hyper-parameters α and β on the Corel5k (Fig5. c) and Pascal07 (Fig5. d) datasets are presented. Both datasets contain 50% avail￾able views and labels, wi…
Figure 6
Figure 6. Figure 6: A random samples feature similarity heat maps across [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Temporal Working Memory: Query-Guided Segment Refinement for Enhanced Multimodal Understanding

    cs.CV 2025-02 conditional novelty 5.0 of 10

    A query-guided frame and audio segment selection module, called TWM, improves nine multimodal models on AVQA, video captioning, and video-text retrieval in the reported tables.

Reference graph

Works this paper leans on

46 extracted references · 42 canonical work pages · cited by 1 Pith paper

  1. [1]

    Spt: Sequence prompt trans- former for interactive image segmentation

    Senlin Cheng and Haopeng Sun. Spt: Sequence prompt trans- former for interactive image segmentation. arXiv preprint arXiv:2412.10224, 2024. 3

    arXiv 2024

  2. [2]

    Ft2tf: First-person statement text-to-talking face genera- tion

    Xingjian Diao, Ming Cheng, Wayner Barrios, and SouYoung Jin. Ft2tf: First-person statement text-to-talking face genera- tion. arXiv preprint arXiv:2312.05430, 2023. 1

    arXiv 2023

  3. [3]

    Learning musi- cal representations for music performance question answering

    Xingjian Diao, Chunhui Zhang, Tingxuan Wu, Ming Cheng, Zhongyu Ouyang, Weiyi Wu, and Jiang Gui. Learning musi- cal representations for music performance question answering. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 2803–2813, 2024. 1

    work page 2024

  4. [4]

    Duygulu, K

    P. Duygulu, K. Barnard, J. F. G. de Freitas, and D. A. Forsyth. Object recognition as machine translation-i : Learning a lexi- con for a fixed image vocabulary. 2002. 6

    work page 2002

  5. [5]

    Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christo- pher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision, 111:98 – 136, 2014. 6

    work page 2014

  6. [6]

    Clough, Henning M ¨uller, and Thomas Deselaers

    Michael Grubinger, Paul D. Clough, Henning M ¨uller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evalua- tion resource for visual information systems. 2006. 6

    work page 2006

  7. [7]

    Contrastive multiview subspace clustering of hyperspectral images based on graph convolutional networks

    Renxiang Guan, Zihao Li, Wenxuan Tu, Jun Wang, Yue Liu, Xianju Li, Chang Tang, and Ruyi Feng. Contrastive multiview subspace clustering of hyperspectral images based on graph convolutional networks. IEEE Transactions on Geoscience and Remote Sensing, 62:1–14, 2024. 1

    work page 2024

  8. [8]

    Spatial-spectral graph contrastive clustering with hard sample mining for hyperspectral images

    Renxiang Guan, Wenxuan Tu, Zihao Li, Hao Yu, Dayu Hu, Yuzeng Chen, Chang Tang, Qiangqiang Yuan, and Xinwang Liu. Spatial-spectral graph contrastive clustering with hard sample mining for hyperspectral images. IEEE Transactions on Geoscience and Remote Sensing, pages 1–16, 2024. 1

    work page 2024

Show all 46 references
  1. [9]

    Girshick

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll’ar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 15979–15988, 2021. 3

    2022

  2. [10]

    Huiskes and Michael S

    Mark J. Huiskes and Michael S. Lew. The mir flickr retrieval evaluation. In Multimedia Information Retrieval, 2008. 6

    2008

  3. [11]

    Rethinking multi-view representation learning via distilled disentangling

    Guanzhou Ke, Bo Wang, Xiaoli Wang, and Shengfeng He. Rethinking multi-view representation learning via distilled disentangling. ArXiv, abs/2403.10897, 2024. 2, 3, 5

    2024 arXiv

  4. [12]

    for two-way multi-label loss

    Takumi Kobayashi. for two-way multi-label loss. 2023. 1

    2023

  5. [13]

    A variational information bottleneck approach to multi-omics data inte- gration

    Changhee Lee and Mihaela van der Schaar. A variational information bottleneck approach to multi-omics data inte- gration. In Arindam Banerjee and Kenji Fukumizu, editors, Proceedings of The 24th International Conference on Artifi- cial Intelligence and Statistics, volume 130 o...

    2021

  6. [14]

    A concise yet effective model for non-aligned incomplete multi-view and missing multi- label learning

    Xiang Li and Songcan Chen. A concise yet effective model for non-aligned incomplete multi-view and missing multi- label learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):5918–5932, 2022. 1, 2, 6

    2022

  7. [15]

    Dual label-guided graph re- finement for multi-view graph clustering

    Yawen Ling, Jianpeng Chen, Yazhou Ren, Xiaorong Pu, Jie Xu, Xiao lan Zhu, and Lifang He. Dual label-guided graph re- finement for multi-view graph clustering. InAAAI Conference on Artificial Intelligence, 2023. 1

    2023

  8. [16]

    Attention-induced embed- ding imputation for incomplete multi-view partial multi-label classification

    Chengliang Liu, Jinlong Jia, Jie Wen, Yabo Liu, Xiaoling Luo, Chao Huang, and Yong Xu. Attention-induced embed- ding imputation for incomplete multi-view partial multi-label classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 1386...

    2024

  9. [17]

    Masked two-channel decoupling framework for incomplete multi-view weak multi-label learn- ing

    Chengliang Liu, Jie Wen, Yabo Liu, Chao Huang, Zhihao Wu, Xiaoling Luo, and Yong Xu. Masked two-channel decoupling framework for incomplete multi-view weak multi-label learn- ing. In Advances in Neural Information Processing Systems, volume 36, 2024. 3, 5, 6, 7

    2024

  10. [18]

    Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classi- fication

    Chengliang Liu, Jie Wen, Xiaoling Luo, Chao Huang, Zhihao Wu, and Yong Xu. Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classi- fication. ArXiv, abs/2303.08358, 2023. 2, 6

    2023 arXiv

  11. [19]

    In- complete multi-view multi-label learning via label-guided masked view- and category-aware transformers

    Chengliang Liu, Jie Wen, Xiaoling Luo, and Yong Xu. In- complete multi-view multi-label learning via label-guided masked view- and category-aware transformers. ArXiv, abs/2303.07180, 2023. 2, 6

    2023 arXiv

  12. [20]

    Low-rank multi-view learning in matrix completion for multi-label image classification

    Meng Liu, Yong Luo, Dacheng Tao, Chao Xu, and Yonggang Wen. Low-rank multi-view learning in matrix completion for multi-label image classification. In Proceedings of the Twenty- Ninth AAAI Conference on Artificial Intelligence, AAAI’15, page 27782784. AAAI Press, 2015. 2

    2015

  13. [21]

    Late fusion incomplete multi-view clustering

    Xinwang Liu, Xinzhong Zhu, Miaomiao Li, Lei Wang, Chang Tang, Jianping Yin, Dinggang Shen, Huaimin Wang, and Wen Gao. Late fusion incomplete multi-view clustering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41:2410–2423, 2019. 6

    2019

  14. [22]

    Multi-scale locality preserving projection for partial multi-view incomplete multi-label learning

    Jiang Long, Qi Zhang, Xiaohuan Lu, Jie Wen, Lian Zhao, and Wulin Xie. Multi-scale locality preserving projection for partial multi-view incomplete multi-label learning. Neural networks : the official journal of the International Neural Network Society, 180:106748, 2024. 2

    2024

  15. [23]

    Task-augmented cross-view imputa- tion network for partial multi-view incomplete multi-label classification

    Xiaohuan Lu, Lian Zhao, Wai Keung Wong, Jie Wen, Jiang Long, and Wulin Xie. Task-augmented cross-view imputa- tion network for partial multi-view incomplete multi-label classification. ArXiv, abs/2409.07931, 2024. 2

    2024 arXiv

  16. [24]

    Rewrite the stars

    Xu Ma, Xiyang Dai, Yue Bai, Yizhou Wang, and Yun Fu. Rewrite the stars. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2024. 5

    2024

  17. [25]

    Expand globally, shrink locally: Discriminant multi-label learning with missing labels

    Zhongchen Ma and Songcan Chen. Expand globally, shrink locally: Discriminant multi-label learning with missing labels. Pattern Recognit., 111:107675, 2020. 6

    2020

  18. [26]

    In- complete multi-view multi-label classification via a dual-level contrastive learning framework, 2024

    Bingyan Nie, Wulin Xie, Jiang Long, and Xiaohuan Lu. In- complete multi-view multi-label classification via a dual-level contrastive learning framework, 2024. 2

    2024

  19. [27]

    Ultra-high resolution segmentation via boundary-enhanced patch-merging transformer

    Haopeng Sun. Ultra-high resolution segmentation via boundary-enhanced patch-merging transformer. arXiv preprint arXiv:2412.10181, 2024. 3

    2024 arXiv

  20. [28]

    Program: Prototype graph model based pseudo-label learning for test-time adaptation

    Haopeng Sun, Lumin Xu, Sheng Jin, Ping Luo, Chen Qian, and Wentao Liu. Program: Prototype graph model based pseudo-label learning for test-time adaptation. In The Twelfth International Conference on Learning Representations, 2024. 3

    2024

  21. [29]

    Lcbm: A multi-view probabilistic model for multi-label classification

    Shiliang Sun and Daoming Zong. Lcbm: A multi-view probabilistic model for multi-label classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(8):2682–2696, 2021. 2

    2021

  22. [30]

    Incomplete multi-view weak-label learn- ing

    Qiaoyu Tan, Guoxian Yu, Carlotta Domeniconi, Jun Wang, and Zi Jian Zhang. Incomplete multi-view weak-label learn- ing. In International Joint Conference on Artificial Intelli- gence, 2018. 2, 6

    2018

  23. [31]

    Sample-level multi-view graph clustering

    Yuze Tan, Yixi Liu, Shudong Huang, Wentao Feng, and Jiancheng Lv. Sample-level multi-view graph clustering. 2023 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 23966–23975, 2023. 1

    2023

  24. [32]

    Knowledge amal- gamation for multi-label classification via label dependency transfer

    Jidapa Thadajarassiri, Thomas Hartvigsen, Walter Gerych, Xiangnan Kong, and Elke Rundensteiner. Knowledge amal- gamation for multi-label classification via label dependency transfer. In Proceedings of the Thirty-Seventh AAAI Confer- ence on Artificial Intelligence and Thirty-F...

    2023

  25. [33]

    The information bottleneck method

    Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. ArXiv, physics/0004057,

    arXiv

  26. [34]

    Luis von Ahn and Laura A. Dabbish. Labeling images with a computer game. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 2004. 6

    2004

  27. [35]

    Auto- weighted multi-view clustering for large-scale data

    Xinhang Wan, Xinwang Liu, Jiyuan Liu, Siwei Wang, Yi Wen, Weixuan Liang, En Zhu, Zhe Liu, and Lu Zhou. Auto- weighted multi-view clustering for large-scale data. In AAAI Conference on Artificial Intelligence, 2023. 1

    2023

  28. [36]

    Rethink- ing minimal sufficient representation in contrastive learning

    Haoqing Wang, Xun Guo, Zhiwei Deng, and Yan Lu. Rethink- ing minimal sufficient representation in contrastive learning. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16020–16029, 2022. 4

    2022

  29. [37]

    Multi- view multi-label learning with view-specific information ex- traction

    Xuan Wu, Qing-Guo Chen, Yao Hu, Dengbao Wang, Xi- aodong Chang, Xiaobo Wang, and Min-Ling Zhang. Multi- view multi-label learning with view-specific information ex- traction. In Proceedings of the 28th International Joint Con- ference on Artificial Intelligence, IJCAI’19, page...

    2019

  30. [38]

    Uncertainty-aware pseudo-labeling and dual graph driven network for incomplete multi-view multi-label classification

    Wulin Xie, Xiaohuan Lu, Yadong Liu, Jiang Long, Bob Zhang, Shuping Zhao, and Jie Wen. Uncertainty-aware pseudo-labeling and dual graph driven network for incomplete multi-view multi-label classification. In ACM Multimedia,

  31. [39]

    Reliable conflictive multi-view learning

    Cai Xu, Jiajun Si, Ziyu Guan, Wei Zhao, Yue Wu, and Xiyue Gao. Reliable conflictive multi-view learning. In AAAI Con- ference on Artificial Intelligence, 2024. 1

    2024

  32. [40]

    Multi-level feature learning for contrastive multi-view clustering

    Jie Xu, Huayi Tang, Yazhou Ren, Liang Peng, Xiao lan Zhu, and Lifang He. Multi-level feature learning for contrastive multi-view clustering. 2022 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 16030– 16039, 2021. 3

    2022

  33. [41]

    Deep partial multi-view learn- ing

    Changqing Zhang, Yajie Cui, Zongbo Han, Joey Tianyi Zhou, Huazhu Fu, and Qinghua Hu. Deep partial multi-view learn- ing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2402–2415, 2022. 3

    2022

  34. [42]

    Latent semantic aware multi- view multi-label classification

    Changqing Zhang, Ziwei Yu, Qinghua Hu, Pengfei Zhu, Xin- wang Liu, and Xiaobo Wang. Latent semantic aware multi- view multi-label classification. In Proceedings of the Thirty- Second AAAI Conference on Artificial Intelligence and Thirti- eth Innovative Applications of Artifici...

    2018

  35. [43]

    Non-aligned multi-view multi-label classification via learn- ing view-specific labels

    Dawei Zhao, Qingwei Gao, Yixiang Lu, and Dong Sun. Non-aligned multi-view multi-label classification via learn- ing view-specific labels. IEEE Transactions on Multimedia, 25:7235–7247, 2023. 2, 6

    2023

  36. [44]

    Consistency and diversity neural net- work multi-view multi-label learning

    Dawei Zhao, Qingwei Gao, Yixiang Lu, Dong Sun, and Yusheng Cheng. Consistency and diversity neural net- work multi-view multi-label learning. Knowl. Based Syst., 218:106841, 2021. 2, 6

    2021

  37. [45]

    Global and local multi-view multi- label learning

    Changming Zhu, Duoqian Miao, Zhe Wang, Rigui Zhou, Lai Wei, and Xiafen Zhang. Global and local multi-view multi- label learning. Neurocomput., 371(C):6777, jan 2020. 2

    2020

  38. [46]

    Kwok, and Zhi-Hua Zhou

    Yue Zhu, James T. Kwok, and Zhi-Hua Zhou. Multi-label learning with global and local label correlation.IEEE Transac- tions on Knowledge and Data Engineering, 30(6):1081–1094,

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.