UniMLR: Modeling Implicit Class Significance for Multi-Label Ranking

Alican Mertan; Emine Dari; Gozde Unal; V. Bugra Yesilkaynak

REVIEW 3 major objections 6 minor 36 references

UniMLR: Modeling Implicit Class Significance for Multi-Label Ranking

T0 review · 3 major / 6 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 UniMLR models each multi-label class's hidden significance as a Gaussian, learned from the ordering of positive labels alone.

desk verdict UniMLR is a worthwhile empirical package—new positive-order loss, a controlled Ranked MNIST benchmark, and honest comparisons—but the 'proportional significance' headline outruns the evidence and the probabilistic derivation has a real independence gap. read the letter →

arxiv 2508.21772 v1 pith:KGR2GS6K submitted 2025-08-29 cs.LG

V. Bugra Yesilkaynak , Emine Dari , Alican Mertan , Gozde Unal This is my paper

classification cs.LG

keywords multi-labelrankinglabelclasssignificancepairwiseGaussianmodelRankedMNISTclassificationdistributionlearning

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

Existing multi-label ranking treats positive labels as equally important and infers rank only from the positive/negative split. This paper argues that the ordering among positive labels is usable signal: it lets a model recover an implicit significance score per label, not just a binary relevance. UniMLR models each label's significance as a Gaussian distribution over a shared score space, so a label is predicted positive when its score is above zero and ranked by the score itself. The paper's central claim is that, given enough data, these learned scores become proportional to the true underlying significance values even though only ranks were observed. To test this it introduces eight synthetic Ranked MNIST datasets with known significance factors and reports improvements in ranking metrics on both synthetic and real multi-label ranking data.

What carries the argument

A per-label Gaussian significance score is the load-bearing device. A network f(x) outputs a mean and variance for every label; prediction is the mean mu, classification is the sign of the Gaussian draw, and ranking is the pairwise probability that one score exceeds another, expressed through the Gaussian Q/erf function. The zero threshold is an implicit classifier, so bipartition and rank order live in the same numerical space and are optimized by one differentiable loss.

What would settle it

Train UniMLR on a Ranked MNIST variant with known per-label scale values, then hold out a split where the same positive set appears with two different positive-rank orders across instances; if the predicted mean scores do not stay monotonically proportional to the true significance values (or if the rank order is not recoverable when the classification and ranking signals are deliberately decorrelated), the central proportionality claim fails.

Watch

Extended reading notes

Core claim

UniMLR treats multi-label ranking as learning a Gaussian significance score s_j ~ N(mu_j, sigma_j^2) for each label of an input. The sign of the score performs classification (positive if s >= 0, negative if s < 0), while the relative scores realize the ranking among positives. The training objective is the negative log of a product: a classification term Q(mu_c, sigma_c) for each label and a ranking term Q(mu_u - mu_v, sqrt(sigma_u^2 + sigma_v^2)) for every ordered pair of labels in the bucket order. The paper's strongest claim is that this single model 'accurately learns a representation of the positive rank order' that is consistent with the ground truth and proportional to the underlying

Load-bearing premise

The model's loss assumes that the positive/negative split and the positive-order ranking are independent given the input, but in UniMLR both are generated by the same Gaussian significance scores; if that independence step cannot be justified, the unified loss is an effective heuristic rather than a principled joint probability model.

Editorial extensions

If this is right

Multi-label ranking systems can extract latent relevance intensities from rank-only annotations, without real-valued supervision.
A single model can output both a label set and a ranked list, with the zero score acting as an automatic threshold rather than a separately tuned hyperparameter.
Pairwise constraints between positive labels meaningfully improve ranking quality (Kendall's tau-b, Spearman's rho, Goodman-Kruskal gamma) while keeping F1 comparable to classification-only methods.
Ranked MNIST provides a controllable benchmark family where the generative significance factors (scale, brightness) are known, enabling cleaner measurement of what a ranking model has learned.
On noisy real-world datasets (NSID, AVDP), strong methods using positive ranks dominate weak methods on ranking metrics.

Reading between the lines

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

Editorial inference: the independence assumption between the positive-set likelihood and the bucket-order likelihood is not implied by the model's own generative story, since both Y and B are deterministic functions of the same Gaussian significance scores; the product-of-likelihoods objective is therefore best read as a heuristic decomposition unless a shared-latent derivation is supplied.
Editorial inference: if the proportionality claim holds, the predicted mean scores could serve as pseudo label-distribution annotations, connecting MLR to label distribution learning without real-valued ground truth.
Editorial inference: the predicted per-label variance is a natural uncertainty signal; a testable extension is to use it for abstention or active learning in ranking tasks.
Editorial inference: the Ranked MNIST generator could be extended to other continuous object attributes (angle, color temperature, blur) to probe which visual factors the model maps onto significance.

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.

Desk Editor's Note

read the letter

Short take: this paper deserves a serious referee but not as-is. The Ranked MNIST datasets and the positive-positive pairwise Gaussian loss are genuinely useful additions to MLR. The strong baselines (LSEP-Strong) are a good idea and the real-data gains over LSEP-Strong on NSID are real (75.4 vs 72.6 tau-b). The variance experiment with small scale differences shows the method picks up subtle ranking signal. Credit where due: they ship code, release the synthetic benchmarks, and report 5-seed error bars on real data.

The soft spots are the ones your reader flags. Section 3.2 assumes Y and B are independent given x, but under their own generative story both are deterministic functions of the same Gaussian significance scores, so the product-of-likelihoods factorization isn't a valid joint model. It works fine as a heuristic loss; they should say so and drop the probabilistic covering.

The proportionality claim is also over-claimed. Table 1 uses Kendall's tau, Spearman's rho, and Goodman-Kruskal gamma—all ordinal. The calibration plots show fitted Gaussians whose means move monotonically with scale, but there's no test that predicted scores are linear in ground-truth significance. Worse, the loss in Eq. (10) is invariant under strictly increasing transforms of the mu's (with threshold shift), so the optimizer has no reason to recover proportional scores. The paper's own Limitations section concedes 'further experimental and theoretical studies' are needed for calibrated significance learning. That's the load-bearing weakness.

Novelty vis-a-vis Zhang et al. [35] and Brinker & Hullermeier [3] is acknowledged but understated. The Gaussian probabilistic treatment with learned per-class variance is new, and the unifying classification+ranking loss is new in this form. That's enough novelty for a workshop or a solid conference paper once the claims are scaled back.

Minor things: the gain over LSEP-Strong on Ranked MNIST Gray-S is tiny (94.23 vs 93.99 tau-b), though it's a saturated task; the real-data differences are more convincing. The A VDP results don't show much gap, which is fine given label noise.

Bottom line: send to peer review. The method is practical, the benchmark is reusable, and the theoretical framing is fixable. A good referee would push for a direct linearity test (e.g., regress predicted scores on ground-truth significance values), re-framing of the loss as a heuristic, and a sharper novelty statement. With those, it would be a solid contribution.

Referee Report

3 major / 6 minor

Summary. The paper introduces UniMLR, a multi-label ranking method that exploits ranking information among positive labels. It models each label's implicit significance as a Gaussian random variable whose mean and variance are predicted from the input, and trains a unified objective with a classification term (whether the significance score is positive or negative) and a ranking term (pairwise comparisons among positive labels). The authors also introduce eight synthetic Ranked MNIST datasets with controlled significance-determining factors, and report experiments on two real datasets and the synthetic datasets. The paper claims that the learned scores are 'proportional to the underlying significance values' and that the method unifies classification and ranking in a principled probabilistic framework.

Significance. If the central claims were fully established, the paper would offer a practical way to extract latent label-significance values from rank-only annotations, which is an attractive goal for recommendation, vision, and decision systems. The proposed synthetic dataset family is a useful controlled testbed for multi-label ranking research, and the idea of explicitly using positive-positive preference constraints is a reasonable extension of existing pairwise ranking losses. However, the proportionality claim is not supported by the experiments and is partly non-identifiable under the proposed loss. The probabilistic derivation also contains an independence inconsistency. The empirical gains over strong baselines are modest and sometimes not statistically grounded. Still, the core loss formulation and the synthetic benchmark are potentially valuable contributions if the claims are appropriately reframed.

major comments (3)

[Section 3.2, Eq. (1); Section 4.2] The factorization P(Y|x) P(B|x) in Eq. (1) is not a valid joint likelihood under the model described in Section 4.2. There, both the positive set Y and the bucket order B are deterministic functions of the same Gaussian significance vector s: Y = {c : s_c >= 0} and B is the order among positive s-values. Thus Y and B are conditionally dependent given x (and s), and the joint likelihood should be an integral over s of an indicator that Y and B match the observed values. The product form in Eq. (10) is therefore an approximation/heuristic, not a 'probabilistic foundation' as claimed. Please either provide a proper generative model or explicitly state that the objective is a surrogate loss.
[Section 4.3 and Section 5.5] The central claim that predicted scores are 'proportional to the underlying significance values' is not supported and is not identifiable from the proposed loss. Since Q(µ, σ) = 1/2[1 - erf(-µ/(σ√2))], the loss in Eq. (10) is invariant under (µ, σ) -> (cµ, cσ) for any c > 0; the ranking term is also invariant under any strictly increasing transformation of the scores. Therefore the optimizer cannot recover the scale of the underlying significance values. The calibration experiment in Section 5.5 only fits Gaussian densities to predicted scores and provides no statistical test of proportionality, and the metrics in Table 1 (τb, Sρ, γ) are rank correlations. The paper's own Limitations section concedes that 'learning calibrated significance values' requires further study. The authors should either prove identifiability under explicit assumptions or weaken the claim to ordinal consistency,
[Table 1 and Section 5.6] The text states that UniMLR 'slightly outperforms LSEP' on Ranked MNIST Gray, but this is not consistently true. On Ranked MNIST Gray-B, LSEP-Strong achieves τb=93.62 and Sρ=97.01 versus UniMLR's 93.38 and 96.65. More importantly, Table 1 reports no error bars or repeated runs for the synthetic datasets, and the differences to LSEP-Strong are often within a small margin (e.g., Ranked MNIST Gray-S τb: 94.23 vs. 93.99). Without standard errors or significance tests, the claimed empirical advantage is not established. Table 3 provides error bars only for NSID and AVDP; the same repeated-run methodology should be applied to the synthetic experiments.

minor comments (6)

[Section 3.2, Eq. (2)] The symbol B is used both for the Bernoulli probability mass function in Eq. (2) and for the bucket order relation in Eq. (3). This overloads the notation; please use a different symbol, e.g., Ber(y_c; ...).
[Section 5.2] Ranked MNIST Color is introduced as a dataset branch, but no experiments on it are reported anywhere. Either present results for this branch or clearly state that it is left for future work and remove it from the abstract/contribution list if not evaluated.
[Tables 1 and 2] The symbol 'S' is used both for the 'Strong' method variant and for the 'Scale' dataset (e.g., Gray-S). This makes the table columns ambiguous. Rename one of the meanings, for example using 'Strong' spelled out or a different abbreviation.
[Abstract and Section 5] The phrase 'statistically demonstrate' in the abstract is stronger than what the experiments support: the quantitative metrics are rank correlations and the calibration experiment is qualitative. Consider softening to 'empirically investigate' or 'provide evidence for'.
[Eq. (10)] The classification and ranking terms are summed with equal weight. This equal weighting is a free parameter and should be discussed or ablated; a simple grid search over the relative weight could be reported.
[Figure 1 caption] The caption lists the three digits as '⟨5, 4, 8⟩' and '⟨3, 8, 9⟩' but the text refers to 'ylow, ymiddle, yhigh' without mapping. Please clarify the correspondence in the caption so the reader can interpret the curves.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: UniMLR's ranking loss is trained on rank labels and evaluated on held-out significance values; the proportionality overclaim is a correctness risk, not a circular reduction.

full rationale

The derivation chain from Eq. (5) through Eq. (10) is self-contained: UniMLR defines predicted significance as Gaussian parameters output by a network, and the loss L_r is a product of pairwise probabilities P(s_u >= s_v) derived directly from the ground-truth bucket order B(i). The classification term L_c is a Q-function threshold. Nothing in the objective is defined in terms of the ground-truth significance values themselves; those values are used only to generate the rank labels in the synthetic data and to evaluate the learned scores afterward. The calibration experiment (Section 5.5) uses a separately generated image set DC and fits Gaussians only for visualization, not as part of training. Hence the rank-order 'prediction' is a standard supervised generalization measure on held-out data, not a circular fit. The abstract's stronger claim that the learned representation is 'proportional to the underlying significance values' is not established by the rank-correlation metrics or qualitative plots, and the paper's Limitations section concedes this ('UniMLR calls for further experimental and theoretical studies on learning calibrated significance values'); however, an unsupported extrapolation is a correctness risk, not a circular reduction. The independence assumption between Y and B in Eq. (1) is internally inconsistent with the Section 4.1 generative story in which both are deterministic functions of the same s_j, but this is a modeling inconsistency rather than a reduction of the output to the input. The self-citations ([10] for the AVDP dataset, [27] for a related distributional ranking loss) are not load-bearing for the paper's central derivation or experimental claims.

Assumptions & free parameters 1 free parameters · 5 assumptions · 1 invented entities

The central claim rests on several independence assumptions that are stated but not derived, including a factorization that conflicts with the paper's own generative story. The learned significance variable is a latent construct without external validation, and the large-dataset proportionality claim is an unproven assertion. These together limit the theoretical foundation of the method.

free parameters (1)

Equal weighting of classification and ranking losses = lambda = 1 (implicit)
Equation (10) sums Lc and Lr with no balancing hyperparameter; this equal weight is chosen by hand and affects the trade-off between classification and ranking performance.

assumptions (5)

domain assumption All label-significance value pairs (y_j, s_j) are conditionally independent given input x.
Stated in Section 3.1. This independence underpins the factorization of the likelihood into per-label and per-pair terms; if significance values for different labels are correlated given the input, the factorization is invalid.
domain assumption Positive label set Y and bucket order B are independent given input x.
Stated in Section 3.2 to allow Equation (1) as a product of two independent likelihoods. This contradicts the paper's own model where both Y and B are deterministic functions of the same significance scores.
domain assumption The bucket order likelihood factorizes into a product over pairwise comparisons.
Used in Equation (3). This is standard in pairwise ranking methods (RPC, LSEP) but is still a modeling assumption; it ignores dependencies among pairwise preferences.
domain assumption Significance values are Gaussian distributed.
Section 4.1 models each significance value as N(mu, sigma^2). The paper says the distribution choice is flexible, but the Gaussian form is a concrete assumption used for the Q-function in the loss.
ad hoc to paper For a large enough dataset, predicted scores become proportional to underlying significance values.
Section 4.2 states this as a claim ('we claim that our predictions will be proportional to the real underlying significance values') without proof or formal conditions. It is a central assertion that is only empirically illustrated.

invented entities (1)

Implicit class significance value s_j
purpose: A latent variable intended to reconcile classification and ranking: whether s_j is above zero determines positive/negative, and the order of s_j values determines the label ranking.
In synthetic Ranked MNIST, significance is operationally defined as digit scale or brightness, but in real datasets it is not independently observed. The paper asserts the model extracts these values without external ground truth, so there is no falsifiable handle outside this work.

how reviews work

0 comments

Cite this review

Pith. "Pith review of UniMLR: Modeling Implicit Class Significance for Multi-Label Ranking." pith.science (2026). https://pith.science/paper/KGR2GS6K

@misc{pith2026250821772,
  author       = {Pith},
  title        = {Pith review of: UniMLR: Modeling Implicit Class Significance for Multi-Label Ranking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KGR2GS6K}},
  note         = {Machine review of arXiv:2508.21772}
}

read the original abstract

Existing multi-label ranking (MLR) frameworks only exploit information deduced from the bipartition of labels into positive and negative sets. Therefore, they do not benefit from ranking among positive labels, which is the novel MLR approach we introduce in this paper. We propose UniMLR, a new MLR paradigm that models implicit class relevance/significance values as probability distributions using the ranking among positive labels, rather than treating them as equally important. This approach unifies ranking and classification tasks associated with MLR. Additionally, we address the challenges of scarcity and annotation bias in MLR datasets by introducing eight synthetic datasets (Ranked MNISTs) generated with varying significance-determining factors, providing an enriched and controllable experimental environment. We statistically demonstrate that our method accurately learns a representation of the positive rank order, which is consistent with the ground truth and proportional to the underlying significance values. Finally, we conduct comprehensive empirical experiments on both real-world and synthetic datasets, demonstrating the value of our proposed framework.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages

[35]

Zhang, Q.-W

M.-L. Zhang, Q.-W. Zhang, J.-P. Fang, Y .-K. Li, and X. Geng. Lever- aging implicit relative labeling-importance information for effective multi-label learning. IEEE Transactions on Knowledge and Data Engi- neering, 33(5):2057–2070, 2021. doi: 10.1109/TKDE.2019.2951561

work page arXiv
[3]

Brinker and E

K. Brinker and E. Hüllermeier. Case-based multilabel ranking. IJ- CAI’07, page 702–707, San Francisco, CA, USA, 2007. Morgan Kauf- mann Publishers Inc

work page 2007
[1]

Alazaidah, F

R. Alazaidah, F. K. Ahmad, and M. F. Mohamad Mohsin. Multi label ranking based on positive pairwise correlations among labels. Interna- tional Arab Journal of Information Technology, 17, 2019

work page 2019
[2]

Brinker and E

K. Brinker and E. Hüllermeier. Case-based multilabel ranking. In IJ- CAI, 2007

work page 2007
[4]

Brinker, J

K. Brinker, J. Fürnkranz, and E. Hüllermeier. A unified model for mul- tilabel classification and ranking. In ECAI, page 489–493, 2006

work page 2006
[5]

S. S. Bucak, P. K. Mallapragada, R. Jin, and A. K. Jain. Efficient multi- label ranking for multi-class learning: Application to object recognition. ICCV, pages 2098–2105, 2009

work page
[6]

Z. Cao, T. Qin, T.-Y . Liu, M.-F. Tsai, and H. Li. Learning to rank: From pairwise approach to listwise approach. In ICML, pages 129–136, 01 2007

work page 2007
[7]

M. Chen, A. Zheng, and K. Weinberger. Fast image tagging. In ICML, pages 1274–1282, 2013

work page 2013

Show all 36 references

[8]

Clare and R

A. Clare and R. D. King. Knowledge discovery in multi-label phenotype data. In L. De Raedt and A. Siebes, editors, Principles of Data Min- ing and Knowledge Discovery, pages 42–53, Berlin, Heidelberg, 2001. Springer Berlin Heidelberg. ISBN 978-3-540-44794-8

2001
[9]

Dembczynski, W

K. Dembczynski, W. Kotłowski, and E. Hüllermeier. Consistent mul- tilabel ranking through univariate loss minimization. In ICML, page 1347–1354, 2012

2012
[10]

Demir, A

G. Demir, A. Çekmi¸ s, V . B. Ye¸ silkaynak, and G. Unal. Detecting vi- sual design principles in art and architecture through deep convolutional neural networks. In Automation in Construction, 2021

2021
[11]

L. Deng. The mnist database of handwritten digit images for machine learning research. In IEEE Signal Process, pages 141–142, 2012

2012
[12]

L. N. Dery. Multi-label ranking: Mining multi-label and label ranking data. ArXiv, abs/2101.00583, 2021

2021 arXiv
[13]

Fagin, R

R. Fagin, R. Kumar, M. Mahdian, D. Sivakumar, and E. Vee. Compar- ing and aggregating rankings with ties. In SIGMOD-SIGACT-SIGART Symposium on Principles of Database Systems, 2004

2004
[14]

Fürnkranz and E

J. Fürnkranz and E. Hüllermeier. Preference learning and ranking by pairwise comparison. In Preference Learning, 2010

2010
[15]

Fürnkranz, E

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

2008
[16]

X. Geng. Label distribution learning. IEEE Transactions on Knowledge and Data Engineering, 28(7):1734–1748, 2016

2016
[17]

X. Geng, R. Zheng, J. Lv, and Y . Zhang. Multilabel ranking with incon- sistent rankers. In PAMI, pages 1–1, 2021

2021
[18]

Y . Gong, Y . Jia, T. Leung, A. Toshev, and S. Ioffe. Deep convolutional ranking for multilabel image annotation. In CoRR, 2014

2014
[19]

Har-Peled, D

S. Har-Peled, D. Roth, and D. Zimak. Constraint classification for mul- ticlass classification and ranking. In NeurIPS, 2002

2002
[20]

Hüllermeier, J

E. Hüllermeier, J. Fürnkranz, W. Cheng, and K. Brinker. Label ranking by learning pairwise preferences. Artificial Intelligence, 172(16-17): 1897–1916, 2008

1916
[21]

Y . H. Jung and A. Tewari. Online boosting algorithms for multi-label ranking. In AISTATS, volume 84, pages 279–287, 2018

2018
[22]

Kanehira and T

A. Kanehira and T. Harada. Multi-label ranking from positive and un- labeled data. In CVPR, pages 5138–5146, 2016

2016
[23]

Y . Li, Y . Song, and J. Luo. Improving pairwise ranking for multi-label image classification. In CVPR, pages 1837–1845, 2017

2017
[24]

Y . Li, K. Liu, R. Satapathy, S. Wang, and E. Cambria. Recent de- velopments in recommender systems: A survey, 2023. URL https: //arxiv.org/abs/2306.12680

2023 arXiv
[25]

Lu and X

Y . Lu and X. Jia. Predicting label distribution from multi-label rank- ing. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems , volume 35, pages 36931–36943. Curran Associates, Inc., 2022

2022
[26]

Y . Lu, W. Li, H. Li, and X. Jia. Predicting label distribution from tie- allowed multi-label ranking. IEEE Transactions on Pattern Analysis and Machine Intelligence , 45(12):15364–15379, 2023. doi: 10.1109/ TPAMI.2023.3300310

2023
[27]

Mertan, Y

A. Mertan, Y . H. Sahin, D. J. Duff, and G. Unal. A new distributional ranking loss with uncertainty: Illustrated in relative depth estimation. In 3DV, pages 1079–1088, 2020

2020
[28]

Sigurbjörnsson and R

B. Sigurbjörnsson and R. van Zwol. Flickr tag recommendation based on collective knowledge. In Proceedings of the 17th International Con- ference on World Wide Web, WWW, page 327–336, 2008

2008
[29]

Toderici, H

G. Toderici, H. B. Aradhye, M. Pasca, L. Sbaiz, and J. Yagnik. Find- ing meaning on youtube: Tag recommendation and category discovery. CVPR, pages 3447–3454, 2010

2010
[30]

Vembu and T

S. Vembu and T. Gärtner. Label ranking algorithms: A survey. In Pref- erence Learning, 2010

2010
[31]

Weston, S

J. Weston, S. Bengio, and N. Usunier. Wsabie: Scaling up to large vocabulary image annotation. In IJCAI, 2011

2011
[32]

G. Wu, C. Li, K. Xu, and J. Zhu. Rethinking and reweighting the uni- variate losses for multi-label ranking: Consistency and generalization. In NeurIPS, 2021

2021
[33]

Xia, T.-Y

F. Xia, T.-Y . Liu, J. Wang, W. Zhang, and H. Li. Listwise approach to learning to rank: Theory and algorithm. In ICML, page 1192–1199, 2008

2008
[34]

Zhang and Z.-H

M.-L. Zhang and Z.-H. Zhou. Multilabel neural networks with appli- cations to functional genomics and text categorization. In IEEE Trans. Knowl. Data Eng., pages 1338–1351, 2006

2006
[36]

Y . Zhou, Y . Liu, J. Yang, X. He, and L. Liu. A taxonomy of label ranking algorithms. J. Comput., 9:557–565, 2014

2014

Pith tools

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

[1] [35]

Zhang, Q.-W

M.-L. Zhang, Q.-W. Zhang, J.-P. Fang, Y .-K. Li, and X. Geng. Lever- aging implicit relative labeling-importance information for effective multi-label learning. IEEE Transactions on Knowledge and Data Engi- neering, 33(5):2057–2070, 2021. doi: 10.1109/TKDE.2019.2951561

work page arXiv

[2] [3]

Brinker and E

K. Brinker and E. Hüllermeier. Case-based multilabel ranking. IJ- CAI’07, page 702–707, San Francisco, CA, USA, 2007. Morgan Kauf- mann Publishers Inc

work page 2007

[3] [1]

Alazaidah, F

R. Alazaidah, F. K. Ahmad, and M. F. Mohamad Mohsin. Multi label ranking based on positive pairwise correlations among labels. Interna- tional Arab Journal of Information Technology, 17, 2019

work page 2019

[4] [2]

Brinker and E

K. Brinker and E. Hüllermeier. Case-based multilabel ranking. In IJ- CAI, 2007

work page 2007

[5] [4]

Brinker, J

K. Brinker, J. Fürnkranz, and E. Hüllermeier. A unified model for mul- tilabel classification and ranking. In ECAI, page 489–493, 2006

work page 2006

[6] [5]

S. S. Bucak, P. K. Mallapragada, R. Jin, and A. K. Jain. Efficient multi- label ranking for multi-class learning: Application to object recognition. ICCV, pages 2098–2105, 2009

work page

[7] [6]

Z. Cao, T. Qin, T.-Y . Liu, M.-F. Tsai, and H. Li. Learning to rank: From pairwise approach to listwise approach. In ICML, pages 129–136, 01 2007

work page 2007

[8] [7]

M. Chen, A. Zheng, and K. Weinberger. Fast image tagging. In ICML, pages 1274–1282, 2013

work page 2013