Pith. sign in

REVIEW 5 major objections 4 minor 45 references

Revisiting Sparsity Constraint Under High-Rank Property in Partial Multi-Label Learning

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Partial multi-label learning should preserve high rank in its predicted labels rather than force low rank.

desk verdict The empirical case for high-rank PML is substantial, but the paper's stated mechanism rests on an invalid nuclear-norm surrogate and a concrete algebraic slip; it needs revision, not rejection. read the letter →

arxiv 2505.20938 v1 pith:B24EQX6C submitted 2025-05-27 cs.LG

Chongjie Si , Yidan Cui , Fuchao Yang , Xiaokang Yang , Wei Shen This is my paper
classification cs.LG
keywords partialmulti-labellearninglabeldisambiguationnoisematrixsparsityconstrainthigh-rankpropertynuclearnormsurrogatelow-rankassumptionclassification
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

Partial multi-label learning (PML) is the setting where every training sample comes with a candidate label set that mixes true labels with noise, and the learner must disambiguate the two. The paper argues that the usual recipe for PML rests on a false conflict: it assumes the noise label matrix is sparse while the true label matrix is low-rank, but real observed and true label matrices are full-rank or nearly full-rank, so the two assumptions cannot plausibly hold together. The authors claim that sparsity and high rank are complementary, because a sparse noise matrix perturbs the singular values of a full-rank matrix very little, so the predicted label matrix keeps the observed matrix's high rank while dropping the noise. They propose Schirn, which adds a negative nuclear-norm term $-\beta\|XW\|_*$ to penalize low-rank predictions, and report that it beats nine baselines on five real-world and six synthetic datasets, winning 92.9% of pairwise comparisons. The payoff, if the claim holds, is a simple correction to a widely used modeling assumption rather than a more complex model.

What carries the argument

The central object is the Schirn objective, $\min_{W,N} \|XW-(Y-N)\|_F^2 + \alpha\|N\|_1 - \beta\|XW\|_* + \lambda\|W\|_F^2$ with $N\in\{0,1\}^{n\times \ell}$ and $N \le Y$. The load-bearing mechanism is the replacement of the rank function by the nuclear norm $\|\cdot\|_*$ (the sum of singular values), so that $-\beta\|XW\|_*$ acts as a convex surrogate for maximizing $\operatorname{rank}(XW)$; a classical singular-value perturbation bound supplies the rank-preservation argument that justifies pairing this with the sparse-noise term $\alpha\|N\|_1$. The optimization is carried by the Augmented Lagrange Multiplier method with a closed-form update for $W$, an ISTA-based shrinkage step for the binary $N$, and singular-value shrinkage for $C=XW$.

What would settle it

Run Schirn with $\beta=0$ and with the nuclear-norm surrogate replaced by an explicit rank-maximizing penalty (for example, maximizing the smallest singular value or flattening the singular-value spectrum) on the same datasets, and compare singular-value spectra at convergence. If the spectra are not flatter under the surrogate than under $\beta=0$, or if the explicit rank penalty reproduces all the gains, then the nuclear-norm surrogate itself is not doing the work the paper assigns it.

Watch

Extended reading notes

Core claim

Schirn's central claim is that in PML, the sparsity constraint on the noise label matrix $N$ and a high-rank constraint on the predicted label matrix $XW$ should be imposed together, replacing the low-rank constraint used by existing methods. Theorem 3.1 states that subtracting a sparse binary matrix $N$ with $\|N\|_0 \le \epsilon$ from a full-rank observed matrix $Y$ leaves a ground-truth matrix $Y_g = Y-N$ with rank at least $\min(n,\ell)-\epsilon$, so sparsity preserves high rank instead of forcing low rank. Based on this, the paper formulates Schirn as the objective in Eq. (5) and solves it by alternating updates for the classifier $W$, the noise matrix $N$, and an auxiliary variable $C=XW$ using shrinkage and singular-value thresholding. The experiments support the claim that the high-rank plus sparsity combination outperforms sparsity-only and sparsity-plus-low-rank methods, and the ablations show that removing either the high-rank term or the sparsity term degrades performance.

Load-bearing premise

The load-bearing premise is that minimizing the negative nuclear norm $-\beta\|XW\|_*$ is a valid way to maximize the rank of the predicted label matrix; if the surrogate does not genuinely promote high rank, the empirical gains attributed to the high-rank mechanism could come from rescaling or noise fitting.

Editorial extensions

If this is right

  • Low-rank regularization is not needed for label disambiguation in PML; on real-world benchmarks it is the wrong inductive bias, because sparse noise leaves the observed label matrix full-rank.
  • A sparse noise matrix does double duty: it removes noisy labels and, by perturbing singular values only slightly, preserves the high-rank structure that carries genuine label information.
  • Ablations show that dropping either the sparsity term or the high-rank term hurts performance, so the gain is attributable to the combination rather than to one term alone.
  • Schirn's predicted matrix rank tracks the ground-truth rank more closely than the observed matrix does, indicating that the method filters noise without collapsing label structure.
  • If the 92.9% pairwise win rate holds up, the sparsity-plus-high-rank objective is a stronger default for PML than the sparsity-plus-low-rank objectives currently in use.

Reading between the lines

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

  • The paper leaves unproven that minimizing $-\beta\|XW\|_*$ actually maximizes $\operatorname{rank}(XW)$; a direct test is to compare against an explicit rank-promoting penalty and inspect the singular-value spectrum at convergence.
  • The perturbation argument bounds singular-value shifts by the perturbation norm, so the conclusion depends on the noise entries being small in magnitude relative to the smallest singular values of $Y$; binary 0/1 noise is sparse but not always small in that sense.
  • Because the method is demonstrated on linear classifiers, the natural testable extension is a deep-network variant where the high-rank term is applied to the penultimate-layer logits; the paper's own limitation section flags this as future work.
  • A stress test for the high-rank assumption would push the noise rate $r$ up until candidate sets become dense; at that point the sparsity assumption fails and the method should degrade, marking the boundary of the regime the paper addresses.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper challenges the common PML assumption that the ground-truth label matrix is low-rank, arguing that sparsity of the label-noise matrix and low-rankness are incompatible because real label matrices are full-rank or nearly so. It proposes Schirn, which adds a sparsity penalty on the noise matrix and a negative nuclear-norm term on the predicted label matrix to encourage high rank, and solves the resulting problem with an ALM scheme. Experiments are reported on five real-world and six synthetic PML datasets against nine baselines, with pairwise t-tests and ablations; the paper claims wins in 92.9% of comparisons.

Significance. If the main mechanism were established, this would be a noteworthy challenge to the low-rank assumption in partial multi-label learning, with potential implications for other weakly supervised multi-label settings. The empirical scope is a real strength: eleven datasets, nine baselines, five metrics, pairwise significance tests, an ablation study, and a supplementary experiment with artificially induced low-rank label matrices. However, the central relaxation from rank maximization to nuclear-norm maximization is not justified, and the optimization contains an algebraic error that directly affects the claimed high-rank behavior. The paper therefore currently provides suggestive empirical evidence rather than a validated mechanism.

major comments (5)
  1. [§3.3, Eqs. (3)–(4)] The replacement of max rank(XW) by max ||XW||_* (equivalently min -β||XW||_*) is not a valid convex relaxation. The nuclear norm is a convex surrogate for rank minimization, not for rank maximization, and it is scale-sensitive: a rank-1 matrix with a large singular value can have a larger nuclear norm than a full-rank matrix with small singular values. In Eq. (5), nothing normalizes the scale of XW, so minimizing -β||XW||_* may enlarge the singular values of a low-rank solution instead of increasing its rank. The paper needs either a proof that this surrogate promotes the desired rank behavior under the specific constraints of Eq. (5), or a controlled experiment (e.g., using a normalized nuclear norm such as ||XW||_*/||XW||_F, or comparing with an explicit rank-maximization baseline) to distinguish rank promotion from simple rescaling.
  2. [§4.3, Eq. (16)] The singular-value update in Eq. (16) is off by a factor of two. The C-subproblem in Eq. (15) has the form min_C (1+µ/2)||C-G||_F^2 - β||C||_*, whose stationary condition for each singular value is (2+µ)(σ_i - g_i) - β = 0, i.e., σ_i = g_i + β/(2+µ), not g_i + 2β/(2+µ). The displayed update therefore over-shifts every singular value by a factor of two. This is not a cosmetic issue: the positive shift makes zero singular values positive in one step, so the rank of C, and eventually of XW, can become full as an artifact of the algorithm rather than as a consequence of a faithful high-rank objective.
  3. [§5.5.2, Table 6] Table 6 is presented as evidence that Schirn preserves high rank, but this is circular. The objective in Eq. (5) explicitly penalizes -||XW||_*, and the C-update in Eq. (16) adds a positive shift to every singular value, so a high-rank output is expected by construction. Additionally, the table does not state how the rank of the continuous prediction matrix P is computed (e.g., numerical tolerance or threshold), which matters for matrices with many small nonzero singular values. The ablation in Table 5 is more informative, but the paper should replace Table 6 with a scale-controlled experiment, for example reporting the singular-value distribution of XW under varying β, or comparing against a baseline that merely rescales a low-rank solution.
  4. [§5.1 and §5.3] The experimental protocol for the three Yeast datasets is not described at the level needed to assess fairness: the paper states that samples without ground-truth labels were removed and that additional random noise was added (r = 3, 5, 10), but it does not report how many samples were removed or how the resulting datasets compare with the original ones used by the baselines. More importantly, α, β, and λ are tuned per dataset over the ranges given in §5.3, yet no inner validation procedure is described. If the hyperparameters were selected using the same test folds whose metrics are later reported, the pairwise win counts at the 0.05 level are optimistically biased. The paper should state the validation protocol and, ideally, report mean performance over the hyperparameter grid or a nested cross-validation.
  5. [Appendix A and Theorem 3.1] The proof of Theorem 3.1 is essentially the rank inequality rank(Y-N) ≥ rank(Y) - rank(N) combined with a bound on rank(N) in terms of the number of nonzero entries. This is a standard perturbation bound and does not establish the paper's stronger claim that 'the sparsity constraint contributes to the high-rank property of the predicted label matrix.' Moreover, the theorem assumes Y is exactly full rank, which is a strong assumption for real data. The authors should either rephrase the claim as a standard stability result under sparse perturbations, with explicit conditions on the number and placement of nonzero entries, or provide a result that connects the sparsity regularization in Eq. (5) to the rank of the recovered label matrix in a way that is not tautological.
minor comments (4)
  1. [Author affiliation] The affiliation line contains a typo: 'Univeristy' should be 'University'.
  2. [Eq. (13)] The composition T_Y(Tsgn(S_{α/L_f}(M))) is not explained in the text; in particular, the behavior when the shrinkage step returns a negative value is not discussed, and the notation Tsgn(a) = 1 if a > 0 and 0 otherwise makes it unclear how negative entries are handled before the thresholding T_Y.
  3. [Algorithm 1] The initialization C = Λ = 1_{n×l} is unusual and unexplained; the text in §4.4 uses Λ←0 implicitly, and the all-ones initialization may affect the first few iterations. Please state the rationale or set them consistently.
  4. [§5.3] The phrase 'β was tuned from [0.01,0.1] with step size 0.01' implies 10 values, but the sensitivity analysis in Figure 2 does not show which metric is plotted or on which dataset; please clarify the figure.

Circularity Check

1 steps flagged · score 4.0 of 10

Table 6's rank-preservation validation is circular by construction; the main comparison against nine baselines is independent, so circularity is partial.

  1. self definitional [Section 5.5.2, Table 6 (with Eq. (5) and Eq. (16))]
    "To evaluate whether Schirn preserves the high-rank property of the predicted label matrix, we compare the rank of the prediction P, denoted r(P), with that of the observed label matrix r(Y) and the ground-truth matrix r(Yg)... Schirn produces a prediction matrix whose rank closely aligns with the ground-truth, while substantially filtering out the noise inherent in Y. In contrast, omitting the high-rank term leads to a noticeable drop in matrix rank."

    The term being validated is the objective being optimized. Eq. (5) contains -β||XW||_*, the explicit 'high-rank' term, and the C-subproblem update Eq. (16), C=U max(0, Σ + 2β/(2+µ)I)V^T, shifts every singular value of C upward by 2β/(2+µ), mechanically raising rank before the thresholding at zero. The Lagrange constraint C=XW then transfers this rank increase to the prediction P. With β=0 the shift disappears, so the observed rank drop is a by-construction consequence of the update. Table 6 therefore does not empirically test whether high rank 'preserves' label structure; it reports that the algorithm's own surrogate was active.

full rationale

The central empirical claim of the paper is the 92.9% pairwise win rate of Schirn against nine PML baselines (Table 4). This comparison is against external methods with their own independently proposed objectives, so it does not reduce to the paper's inputs. The main theoretical step, Theorem 3.1, is a standard rank inequality (rank(Y-N) ≥ rank(Y) - rank(N)) applied to a full-rank Y and sparse N; although elementary, it is not circular because it is a statement about matrices rather than a fitted prediction. The paper does cite the first author's prior work [20] for the full-rank observation, but Table 1 computes ranks directly from datasets, so that self-citation is not load-bearing. The one genuinely circular element is the Section 5.5.2 validation: Schirn enforces 'high rank' by explicitly maximizing a surrogate (via -β||XW||_* and the additive singular-value shift in Eq. (16)), and then Table 6 reports that the output has high rank when β>0 and lower rank when β=0. That is a check that the objective did what it was set to do, not an independent demonstration of the high-rank mechanism. Separately, the replacement of max rank(XW) by max ||XW||_* in Eqs. (3)-(4) is a correctness/soundness concern (maximizing nuclear norm is scale-sensitive and is not a valid surrogate for rank maximization), but it is not a circularity; it weakens the mechanism rather than making the prediction equal to the input. Overall, because the headline performance comparison is externally grounded, the circularity score is 4 rather than higher.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central derivation rests on an elementary rank inequality and on an unproven surrogate for rank maximization. The three free parameters are tuned per dataset, and the full-rank label assumption is domain-specific and only partially verified.

free parameters (3)
  • α = grid [0.1,2] step 0.1
    Sparsity weight on ||N||1, tuned per dataset (§5.3).
  • β = grid [0.01,0.1] step 0.01
    High-rank term weight on -||XW||_*, tuned per dataset (§5.3).
  • λ = {0.1,10,100,250,1000}
    Ridge regularization on W, tuned per dataset (§5.3).
assumptions (4)
  • standard math rank(A-B) ≥ rank(A) - rank(B) for any matrices A, B
    Used in Appendix A to prove Theorem 3.1 (Eq. 20).
  • domain assumption The observed label matrix Y is full-rank and the noise matrix N is sparse with ||N||0 < n, l
    Assumed in Theorem 3.1 and in §1; verified only on a few datasets.
  • ad hoc to paper Minimizing -β||XW||_* is an acceptable surrogate for maximizing rank(XW)
    Introduced in §3.3 Eqs. (3)-(4) without proof; the standard nuclear-norm surrogate applies to rank minimization, not maximization.
  • domain assumption The true label matrix is full-rank or near full-rank in real-world PML
    Core motivation in §1, supported by Table 1 but depends on the filtered and altered datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Revisiting Sparsity Constraint Under High-Rank Property in Partial Multi-Label Learning." pith.science (2026). https://pith.science/paper/B24EQX6C

@misc{pith2026250520938,
  author       = {Pith},
  title        = {Pith review of: Revisiting Sparsity Constraint Under High-Rank Property in Partial Multi-Label Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B24EQX6C}},
  note         = {Machine review of arXiv:2505.20938}
}
read the original abstract

Partial Multi-Label Learning (PML) extends the multi-label learning paradigm to scenarios where each sample is associated with a candidate label set containing both ground-truth labels and noisy labels. Existing PML methods commonly rely on two assumptions: sparsity of the noise label matrix and low-rankness of the ground-truth label matrix. However, these assumptions are inherently conflicting and impractical for real-world scenarios, where the true label matrix is typically full-rank or close to full-rank. To address these limitations, we demonstrate that the sparsity constraint contributes to the high-rank property of the predicted label matrix. Based on this, we propose a novel method Schirn, which introduces a sparsity constraint on the noise label matrix while enforcing a high-rank property on the predicted label matrix. Extensive experiments demonstrate the superior performance of Schirn compared to state-of-the-art methods, validating its effectiveness in tackling real-world PML challenges.

Figures

Figures reproduced from arXiv: 2505.20938 by the authors.

Figure 1
Figure 1. An example of partial multi-label learning. Formally, let X = R d represent the d-dimensional feature space, and let Y = {1, 2, . . . , l} denote the label space with l possible labels. Consider a partial multi-label training set D = {(xi , Ci)| 1 ≤ i ≤ n}, where n is the number of training samples. Here, xi ∈ X is the d-dimensional feature vector for the i-th instance, and Ci ⊆ Y is its candidate label set. The goa… view at source ↗
Figure 2
Figure 2. Sensitivity analysis of Schirn. 6 Conclusion In this paper, we propose Schirn, a novel approach to Partial Multi-Label Learning (PML) that overcomes the limitations of existing methods, which rely on conflicting assumptions of sparsity in noisy label matrices and low-rankness in true label matrices. We theoretically show that the rank of a full-rank matrix remains high under sparse perturbations, forming the foundat… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 39 canonical work pages

  1. [1]

    A fast iterative shrinkage-thresholding algorithm for linear inverse problems.SIAM journal on imaging sciences, 2(1):183–202, 2009

    Amir Beck and Marc Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems.SIAM journal on imaging sciences, 2(1):183–202, 2009

    2009

  2. [2]

    Learning multi-label scene classification.Pattern recognition, 37(9):1757–1771, 2004

    Matthew R Boutell, Jiebo Luo, Xipeng Shen, and Christopher M Brown. Learning multi-label scene classification.Pattern recognition, 37(9):1757–1771, 2004

    work page 2004

  3. [3]

    A singular value thresholding algorithm for matrix completion.SIAM Journal on optimization, 20(4):1956–1982, 2010

    Jian-Feng Cai, Emmanuel J Candès, and Zuowei Shen. A singular value thresholding algorithm for matrix completion.SIAM Journal on optimization, 20(4):1956–1982, 2010

    work page 1956

  4. [4]

    Exact matrix completion via convex optimization

    Emmanuel Candes and Benjamin Recht. Exact matrix completion via convex optimization. Communications of the ACM, 55(6):111–119, 2012

    2012

  5. [5]

    Decoding by linear programming.IEEE transactions on information theory, 51(12):4203–4215, 2005

    Emmanuel J Candes and Terence Tao. Decoding by linear programming.IEEE transactions on information theory, 51(12):4203–4215, 2005

    2005

  6. [6]

    iatc-misf: a multi-label clas- sifier for predicting the classes of anatomical therapeutic chemicals.Bioinformatics, 33(3):341– 346, 2017

    Xiang Cheng, Shu-Guang Zhao, Xuan Xiao, and Kuo-Chen Chou. iatc-misf: a multi-label clas- sifier for predicting the classes of anatomical therapeutic chemicals.Bioinformatics, 33(3):341– 346, 2017

    work page 2017

  7. [7]

    J Fürnkranz, E Hüllermeier, E Loza?Mencía, and K. Brinker. Multilabel classification via calibrated label ranking.Machine Learning, 73(2):133–153, 2008

    work page 2008

  8. [8]

    A tutorial on multilabel learning.ACM Computing Surveys (CSUR), 47(3):1–38, 2015

    Eva Gibaja and Sebastián Ventura. A tutorial on multilabel learning.ACM Computing Surveys (CSUR), 47(3):1–38, 2015

    work page 2015

Show all 45 references
  1. [9]

    Partial multi-label learning with probabilistic graphical disambiguation.Advances in Neural Information Processing Systems, 36:1339–1351, 2023

    Jun-Yi Hang and Min-Ling Zhang. Partial multi-label learning with probabilistic graphical disambiguation.Advances in Neural Information Processing Systems, 36:1339–1351, 2023

    2023

  2. [10]

    Learning label specific features for multi-label classification

    Jun Huang, Guorong Li, Qingming Huang, and Xindong Wu. Learning label specific features for multi-label classification. In2015 IEEE International conference on data mining, pages 181–190. IEEE, 2015

    2015

  3. [11]

    Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection a...

    1956

  4. [12]

    Instance-aware hashing for multi-label image retrieval.IEEE Transactions on Image Processing, 25(6):2469–2479, 2016

    Hanjiang Lai, Pan Yan, Xiangbo Shu, Yunchao Wei, and Shuicheng Yan. Instance-aware hashing for multi-label image retrieval.IEEE Transactions on Image Processing, 25(6):2469–2479, 2016

    2016

  5. [13]

    The augmented lagrange multiplier method for exact recovery of corrupted low-rank matrices.arXiv preprint arXiv:1009.5055, 2010

    Zhouchen Lin, Minming Chen, and Yi Ma. The augmented lagrange multiplier method for exact recovery of corrupted low-rank matrices.arXiv preprint arXiv:1009.5055, 2010

    2010 arXiv

  6. [14]

    Prior knowledge regularized self-representation model for partial multilabel learning.IEEE Transactions on Cybernetics, 53(3):1618–1628, 2021

    Gengyu Lyu, Songhe Feng, Yi Jin, Tao Wang, Congyan Lang, and Yidong Li. Prior knowledge regularized self-representation model for partial multilabel learning.IEEE Transactions on Cybernetics, 53(3):1618–1628, 2021

    2021

  7. [15]

    Partial multi-label learning via probabilistic graph matching mechanism

    Gengyu Lyu, Songhe Feng, and Yidong Li. Partial multi-label learning via probabilistic graph matching mechanism. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 105–113, 2020

    2020

  8. [16]

    Expand globally, shrink locally: Discriminant multi-label learning with missing labels.Pattern Recognition, 111:107675, 2021

    Zhongchen Ma and Songcan Chen. Expand globally, shrink locally: Discriminant multi-label learning with missing labels.Pattern Recognition, 111:107675, 2021

    2021

  9. [17]

    Correl- ative multi-label video annotation

    Guo-Jun Qi, Xian-Sheng Hua, Yong Rui, Jinhui Tang, Tao Mei, and Hong-Jiang Zhang. Correl- ative multi-label video annotation. InProceedings of the 15th ACM international conference on Multimedia, pages 17–26, 2007

    2007

  10. [18]

    J. Read, B. Pfahringer, G. Holmes, and E. Frank. Classifier chains for multi-label classification. InProceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases: Part II, 2009. 10

    2009

  11. [19]

    Augmented lagrange multiplier functions and duality in nonconvex programming.SIAM Journal on Control, 12(2):268–285, 1974

    R Tyrrell Rockafellar. Augmented lagrange multiplier functions and duality in nonconvex programming.SIAM Journal on Control, 12(2):268–285, 1974

    1974

  12. [20]

    Multi- label classification with high-rank and high-order label correlations.IEEE Transactions on Knowledge and Data Engineering, 36(8):4076–4088, 2023

    Chongjie Si, Yuheng Jia, Ran Wang, Min-Ling Zhang, Yanghe Feng, and Chongxiao Qu. Multi- label classification with high-rank and high-order label correlations.IEEE Transactions on Knowledge and Data Engineering, 36(8):4076–4088, 2023

    2023

  13. [21]

    Global-local label correlation for partial multi-label learning.IEEE Transactions on Multimedia, 24:581–593, 2021

    Lijuan Sun, Songhe Feng, Jun Liu, Gengyu Lyu, and Congyan Lang. Global-local label correlation for partial multi-label learning.IEEE Transactions on Multimedia, 24:581–593, 2021

    2021

  14. [22]

    Partial multi-label learning with noisy side information.Knowledge and Information Systems, 63:541–564, 2021

    Lijuan Sun, Songhe Feng, Gengyu Lyu, Hua Zhang, and Guojun Dai. Partial multi-label learning with noisy side information.Knowledge and Information Systems, 63:541–564, 2021

    2021

  15. [23]

    Partial multi-label learning by low-rank and sparse decomposition

    Lijuan Sun, Songhe Feng, Tao Wang, Congyan Lang, and Yi Jin. Partial multi-label learning by low-rank and sparse decomposition. InProceedings of the AAAI conference on artificial intelligence, volume 33, pages 5016–5023, 2019

    2019

  16. [24]

    Large scale multi-label classification via metalabeler

    Lei Tang, Suju Rajan, and Vijay K Narayanan. Large scale multi-label classification via metalabeler. InProceedings of the 18th international conference on World wide web, pages 211–220, 2009

    2009

  17. [25]

    Random k-labelsets for multilabel classification.IEEE Transactions on Knowledge and Data Engineering, 23(7):1079–1089, 2011

    Grigorios Tsoumakas, Ioannis Katakis, and Ioannis Vlahavas. Random k-labelsets for multilabel classification.IEEE Transactions on Knowledge and Data Engineering, 23(7):1079–1089, 2011

    2011

  18. [26]

    Discriminative and correlative partial multi-label learning

    Haobo Wang, Weiwei Liu, Yang Zhao, Chen Zhang, Tianlei Hu, and Gang Chen. Discriminative and correlative partial multi-label learning. InIJCAI, pages 3691–3697, 2019

    2019

  19. [27]

    Deep partial multi-label learning with graph disambiguation.arXiv preprint arXiv:2305.05882, 2023

    Haobo Wang, Shisong Yang, Gengyu Lyu, Weiwei Liu, Tianlei Hu, Ke Chen, Songhe Feng, and Gang Chen. Deep partial multi-label learning with graph disambiguation.arXiv preprint arXiv:2305.05882, 2023

    2023 arXiv

  20. [28]

    Partial multi-label feature selection

    Jing Wang, Peipei Li, and Kui Yu. Partial multi-label feature selection. In2022 International Joint Conference on Neural Networks (IJCNN), pages 1–9. IEEE, 2022

    2022

  21. [29]

    Ml-lrc: Low-rank-constraint-based multi- label learning with label noise

    Xiaoying Wang, Jun Xie, Lu Yu, and Xingliu Tao. Ml-lrc: Low-rank-constraint-based multi- label learning with label noise. In2020 IEEE 4th Information Technology, Networking, Elec- tronic and Automation Control Conference (ITNEC), volume 1, pages 129–136, 2020

    2020

  22. [30]

    Perturbation bounds in connection with singular value decomposition.BIT Numerical Mathematics, 12:99–111, 1972

    Per-Åke Wedin. Perturbation bounds in connection with singular value decomposition.BIT Numerical Mathematics, 12:99–111, 1972

    1972

  23. [31]

    Partial multi-label learning

    Ming-Kun Xie and Sheng-Jun Huang. Partial multi-label learning. InProceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  24. [32]

    Partial multi-label learning with noisy label identification

    Ming-Kun Xie and Sheng-Jun Huang. Partial multi-label learning with noisy label identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3676–3687, 2021

    2021

  25. [33]

    Partial multi-label learning with label distribution

    Ning Xu, Yun-Peng Liu, and Xin Geng. Partial multi-label learning with label distribution. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 6510–6517, 2020

    2020

  26. [34]

    Noisy label removal for partial multi-label learning

    Fuchao Yang, Yuheng Jia, Hui Liu, Yongqiang Dong, and Junhui Hou. Noisy label removal for partial multi-label learning. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3724–3735, 2024

    2024

  27. [35]

    Feature-induced partial multi-label learning

    Guoxian Yu, Xia Chen, Carlotta Domeniconi, Jun Wang, Zhao Li, Zili Zhang, and Xindong Wu. Feature-induced partial multi-label learning. In2018 IEEE international conference on data mining (ICDM), pages 1398–1403. IEEE, 2018

    2018

  28. [36]

    Partial multi-label learning using label compression

    Tingting Yu, Guoxian Yu, Jun Wang, Carlotta Domeniconi, and Xiangliang Zhang. Partial multi-label learning using label compression. In2020 IEEE International Conference on Data Mining (ICDM), pages 761–770. IEEE, 2020. 11

    2020

  29. [37]

    Partial multi-label learning via credible label elicitation

    Min-Ling Zhang and Jun-Peng Fang. Partial multi-label learning via credible label elicitation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3587–3599, 2020

    2020

  30. [38]

    Binary relevance for multi-label learning: an overview.Frontiers of Computer Science, 12:191–202, 2018

    Min-Ling Zhang, Yu-Kun Li, Xu-Ying Liu, and Xin Geng. Binary relevance for multi-label learning: an overview.Frontiers of Computer Science, 12:191–202, 2018

    2018

  31. [39]

    Lift: Multi-label learning with label-specific features.IEEE transactions on pattern analysis and machine intelligence, 37(1):107–120, 2014

    Min-Ling Zhang and Lei Wu. Lift: Multi-label learning with label-specific features.IEEE transactions on pattern analysis and machine intelligence, 37(1):107–120, 2014

    2014

  32. [40]

    Multilabel neural networks with applications to functional genomics and text categorization.IEEE transactions on Knowledge and Data Engineering, 18(10):1338–1351, 2006

    Min-Ling Zhang and Zhi-Hua Zhou. Multilabel neural networks with applications to functional genomics and text categorization.IEEE transactions on Knowledge and Data Engineering, 18(10):1338–1351, 2006

    2006

  33. [41]

    Ml-knn: A lazy learning approach to multi-label learning

    Min-Ling Zhang and Zhi-Hua Zhou. Ml-knn: A lazy learning approach to multi-label learning. Pattern recognition, 40(7):2038–2048, 2007

    2007

  34. [42]

    A review on multi-label learning algorithms.IEEE transactions on knowledge and data engineering, 26(8):1819–1837, 2013

    Min-Ling Zhang and Zhi-Hua Zhou. A review on multi-label learning algorithms.IEEE transactions on knowledge and data engineering, 26(8):1819–1837, 2013

    2013

  35. [43]

    Partial multi-label learning based on sparse asymmetric label correlations.Knowledge-Based Systems, 245:108601, 2022

    Peng Zhao, Shiyi Zhao, Xuyang Zhao, Huiting Liu, and Xia Ji. Partial multi-label learning based on sparse asymmetric label correlations.Knowledge-Based Systems, 245:108601, 2022

    2022

  36. [44]

    Kwok, and Zhi-Hua Zhou

    Yue Zhu, James T. Kwok, and Zhi-Hua Zhou. Multi-label learning with global and local label correlation.IEEE Transactions on Knowledge and Data Engineering, 30(6):1081–1094, 2018

    2018

  37. [45]

    Non- negative low rank and sparse graph for semi-supervised learning

    Liansheng Zhuang, Haoyuan Gao, Zhouchen Lin, Yi Ma, Xin Zhang, and Nenghai Yu. Non- negative low rank and sparse graph for semi-supervised learning. In2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2328–2335. IEEE, 2012. 12 A Proof of Theorem 3.1 We her...

    2012

Pith tools

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