Graph Random Walk with Feature-Label Space Alignment: A Multi-Label Feature Selection Method

Hanlin Pan; Jun Gao; Kunpeng Liu; Qingqi Han; Wanfu Gao

REVIEW 2 major objections 6 minor 43 references

Graph Random Walk with Feature-Label Space Alignment: A Multi-Label Feature Selection Method

T0 review · 2 major / 6 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 A random walk over the feature-label graph, combined with shared low-rank factorization, captures indirect feature-label associations and improves multi-label feature selection.

desk verdict The paper's central claim about capturing indirect associations is undercut by its own update rule: Eq. (8) only reweights direct mutual information, so the random walk does not propagate associations through intermediate nodes. read the letter →

arxiv 2505.23228 v1 pith:SE5LM3C6 submitted 2025-05-29 cs.LG

Wanfu Gao , Jun Gao , Qingqi Han , Hanlin Pan , Kunpeng Liu This is my paper

classification cs.LG

keywords multi-labelfeatureselectionrandomwalkfeature-labelcompositegraphmutualinformationlow-rankmatrixfactorizationspacealignmentmanifoldlearning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes GRW-SCMF, a multi-label feature selection method that ranks features by combining a random walk over a feature-label graph with a shared low-rank factorization of the feature and label matrices. The central idea is that direct mutual information between features and labels misses indirect associations that travel through intermediate features and labels, and the random walk is designed to expose those indirect paths. The method then forces the low-rank product $Q^\top B$ to approximate the random-walk-derived association matrix, and adds a term that aligns feature and label mappings in a common low-dimensional space. On seven benchmark datasets the paper reports that this design beats seven comparison methods on Micro-F1, Macro-F1, hamming loss, and zero-one loss, with one exception on a single metric.

What carries the argument

The load-bearing object is the Random Walk Mutual Information matrix (RWMI), built from a walk on the composite graph $G=(V,E)$ with feature vertices $V_f$, label vertices $V_l$, and three edge types: feature-feature, label-label, and feature-label. Adjacency matrices for feature and label graphs use Gaussian kernels, while the feature-label connection matrix is the mutual information $MI$. A walker moves between feature and label nodes with jump probability $p_{\text{jump}}$, and each feature-label pair encountered at distance $d(f,l)$ updates the weight $RW(f,l) += \text{decay factor}^{d(f,l)} \cdot MI(f,l)$. The resulting normalized RWMI matrix is inserted into the objective as $\|R_w - Q^\top B\|_F^2$, so the random walk's indirect associations are transferred into the ranking scores $\|(Q^\top B)_i\|_2$.

What would settle it

Keep all components fixed but replace RWMI with either the direct mutual information matrix alone or a row-shuffled random matrix; if classification performance does not drop on the reported datasets, the random-walk indirect associations are not what drives the result.

Watch

Extended reading notes

Core claim

The paper claims that random walks on a composite graph whose vertices are features and labels, with edges weighted by Gaussian kernel similarities and mutual information, produce a feature-label association matrix called RWMI that encodes both direct and implicit indirect associations. Embedding this matrix into the objective as $\|R_w - Q^\top B\|_F^2$ makes the shared low-dimensional representation carry high-order relationship information that linear non-negative matrix factorization alone would lose. The alignment term $\|XQ^\top - YB^\top\|_F^2$ keeps similar samples close in the shared space, and feature importance is scored by the row norms of $Q^\top B$. The reported experiments support the claim that the combined model outperforms the seven baselines in almost all configurations.

Load-bearing premise

The method assumes that the random-walk-derived RWMI matrix genuinely captures useful indirect feature-label associations, so that forcing $Q^\top B$ to approximate it improves feature ranking rather than injecting noise.

Editorial extensions

If this is right

Features that connect to labels only through chains of other features or labels receive nontrivial scores from RWMI, so the ranking can promote features that direct mutual information would miss.
Because the random-walk term is combined with shared low-rank reconstruction, the selected feature subset is chosen in a space that respects both feature-label association and label-label correlation.
The space-alignment term means features and labels that describe similar samples are mapped close together, which should carry over to classifiers such as SVM, 3NN, and MLkNN used in the evaluation.
If the reported comparisons hold, GRW-SCMF offers a general ranking procedure for multi-label data across web text, music, image, and biology domains.

Reading between the lines

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

The RWMI update is essentially a graph-diffusion heuristic; an implicit corollary is that replacing the finite random walk with a closed-form diffusion or personalized PageRank over the same composite graph could yield the same indirect associations without tuning walk length or decay factor.
The alignment term resembles cross-modal subspace alignment, so the method may transfer to other paired-data settings, such as aligning image and text embeddings, where a low-rank shared space is desired.
A testable prediction is that the benefit of the RWMI term grows with label-set complexity: datasets with many distinct label combinations should show larger gains over direct-MI baselines than datasets with few labels.

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

The empirical work here is solid and the package is honest: code released, seven datasets, ablations, parameter sweeps. So this is a useful heuristic to know about. The problem is the story the paper tells about why it works. The central novelty is supposed to be the capture of implicit indirect associations between features and labels via random walk. But Eq. (8) updates RW(f,l) by decay_factor^d * MI(f,l). Every increment is multiplied by the direct MI of the endpoint pair. If MI(f,l)=0, the increment is zero, and after normalization the value stays zero. A path f1-f2-l3 contributes decay^2 * MI(f1,l3), not any function of MI(f1,f2) and MI(f2,l3). So RWMI is just a distance-decayed reweighting of the direct MI matrix. The walk changes which pairs are seen and how often, but it does not transmit association along edges. The phrase 'implicit indirect relationships... transmitted through intermediate nodes' is therefore not realized by the algorithm as written. This is not a minor gap; it is the stated motivation for adding the ||Rw - Q^T B|| term. Other soft spots are secondary by comparison. The claim of reaching a global optimum for a non-convex problem is unsupported; the multiplicative updates are standard NMF style and at best guarantee a stationary point. The convergence analysis is just a figure, not a proof. Baseline comparisons are missing significance tests, and the parameter range is wide enough that Bayesian optimization could be fitting per dataset. What holds up: the method is coherent as a distance-decayed MI weighting combined with NMF-based low-rank alignment. That combination appears to be new, and the gains over baselines are consistent across datasets and metrics, even if incremental. The ablation shows both components contribute. As a heuristic, it likely works. But the paper overstates the mechanism, and a reader who accepts the indirect-association story will be misled. This paper deserves a serious referee, because the empirical study is substantial and a good referee would catch the Eq. (8) problem and force the authors to reframe the contribution. Without that, I would not cite it as evidence for indirect-association modeling. The reading group might enjoy deconstructing the mismatch between the narrative and the equations.

Referee Report

2 major / 6 minor

Summary. The paper proposes GRW-SCMF, a multi-label feature selection method. It constructs a composite graph over features and labels, with Gaussian-kernel feature-feature and label-label edges and mutual-information feature-label edges, and runs random walks to produce a matrix Rw via the RWMI update rule. This matrix is then used in a shared low-rank factorization objective that also contains a feature-label space-alignment term and an l2,1 row-sparsity regularizer. Features are ranked by the row norms of Q^T B, where Q and B are the factor coefficients for features and labels. The method is evaluated on seven MULAN datasets against seven baselines using SVM, 3NN, and MLkNN classifiers, with ablations, parameter sensitivity studies, and convergence plots.

Significance. If the proposed mechanism worked as stated, the paper would make a useful contribution: integrating multi-hop random-walk information on a feature-label graph with a shared low-dimensional factorization and a space-alignment term is a sensible design direction for multi-label feature selection. The paper also releases code and reports experiments across several domains and metrics, with ablations and convergence diagnostics, which is a strength. However, the central methodological claim that Eq. (8) captures implicit indirect associations is not supported by the update rule as written: the random walk only reweights direct pairwise mutual information and cannot propagate associations through intermediate nodes. Consequently, the interpretation of the Rw term in the objective is unjustified, and the current empirical gains cannot be attributed to the hypothesized indirect-association mechanism. The results may still be valid as a reweighted-MI method, but the paper needs substantive revision to substantiate or restate its main claim.

major comments (2)

[Section 3, 'Feature-Label Relationship Update Rule', Eq. (8)] Equation (8) updates RW(f,l) by adding decay_factor^{d(f,l)} * MI(f,l). Since every increment is multiplied by the direct feature-label mutual information MI(f,l), any pair with MI(f,l)=0 receives only zero contributions from every walk and remains zero after normalization. Thus no association is ever transmitted from one edge to another through intermediate nodes: the stated mechanism of 'implicit indirect relationships ... transmitted through intermediate nodes' is not realized by this update. The matrix Rw is, as written, a distance-decayed reweighting of the direct MI matrix, where the random walk only determines which feature-label pairs are visited and at what separation. Because the objective term ||Rw - Q^T B|| in Eq. (10) is justified by the supposedly captured indirect associations, the central claim of the paper is unsupported. The authors should either modify the update rule to propagate information along paths (for example, using products of transition probabilities or accumulated MI along the walk) or substantially restate the contribution.
[Section 4, 'Optimization scheme', Eqs. (14) and (20)-(22)] The paper claims that the proposed alternating-multiplier-based relaxation update method is designed to 'achieve the global optimum of the objective function.' This claim is unsupported. The objective in Eq. (13) is non-convex in the variables (V,Q,B) because of the product terms VQ and Q^T B, and the multiplicative updates in Eqs. (20)-(22) are standard NMF-style update rules that can at best be expected to converge to a stationary point or a local minimum under suitable conditions, not to a global optimum. No proof of global optimality is provided, and the statement appears to conflict with the paper's own acknowledgment earlier in the section that the objective is non-convex. Please remove or qualify the global-optimality claim, or supply a rigorous proof if one is intended.

minor comments (6)

[Abstract] The abstract mentions experiments on 'seven benchmark datasets and three representative datasets,' but the experimental section describes only seven datasets total; the meaning of 'three representative datasets' should be clarified or the phrasing corrected.
[Section 5, 'Evaluating Methods'] The description says the top 20% of ranked features are selected 'with a step size of 1%.' If the evaluation takes the best performance over this range of feature-subset sizes, that should be stated explicitly, because selecting the best test-set performance over subset sizes can inflate reported results.
[Section 5, 'Ablation Study', Table 4] Table 4 reports results on only three of the seven datasets and does not include standard deviations, making it difficult to assess the significance of the differences; additionally, the row labels should clarify whether the first two rows are single-component variants of the full method.
[Section 5, 'Experimental Results'] The claim of superiority over all baselines is not accompanied by any statistical significance tests; for several entries, such as Yeast Micro-F1 in Table 2 (0.5900±0.028 vs. 0.5677±0.038), the differences are within one standard deviation, so the strength of the claim should be moderated or supported by paired tests.
[Section 3, 'Objective Function', Eq. (12)] The text states that the Frobenius norm of V 'imposes a sparsity constraint,' but the Frobenius norm encourages small entries, not sparsity; sparsity in this objective comes from the l2,1-norm on Q^T B. This wording should be corrected.
[Section 5, 'Convergence'] The variable z_t used in the stopping criterion is not defined; the convergence discussion should define the objective value or another monitored quantity so that the reported curves are reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

Derivation is self-contained; no circularity found.

full rationale

The paper's derivation chain is data-driven rather than circular. The random walk matrix Rw is constructed from mutual information and graph transition probabilities via Eq. (8), and then the objective function (Eq. 13) fits low-dimensional factors Q and B so that Q^T B approximates Rw. Feature importance is finally computed by the row norms of Q^T B. Nothing in this chain derives the target result from itself: the ranking is a direct algebraic function of the learned factors, and the factors are fitted to the precomputed Rw matrix. The self-citations in the reference list, such as [Gao et al., 2023], are used as comparison baselines or as related work, not as load-bearing justifications for the central mechanism; therefore they do not constitute circularity. The skeptical observation about Eq. (8) is a correctness concern, not a circularity one: because every update is multiplied by MI(f,l), pairs with zero mutual information remain zero after normalization, so the update rule cannot literally create new indirect associations through intermediate nodes. That concern challenges whether the central claim is supported as written, but it does not amount to the paper's output being equivalent to its input by construction. The empirical evaluation is external and the method is fully specified, so the paper merits a circularity score of 0.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The method relies on several free hyperparameters and heuristic assumptions, especially the random walk weighting and the alignment of feature and label spaces. The performance of the method is therefore not a parameter-free derivation, and its success depends on tuning choices that are not fully disclosed.

free parameters (7)

alpha, beta, gamma, delta, epsilon (regularization weights) = not reported
Tuned via Bayesian optimization on each dataset; search range {0.01, 0.1, 0.3, 0.5, 0.7, 0.9, 1.0} (Section 5, Parameter Selection).
sigma (Gaussian kernel width) = not reported
Used in adjacency matrices Afeatures and Alabels; the selection process for sigma is not described in the paper.
n walks = not reported
Number of random walks; Bayesian optimization over {100, 1000, 10000}.
walk length = not reported
Number of steps per random walk; optimized over {10, 20, 30}.
jump prob = not reported
Probability of switching between feature and label nodes; optimized over {0.1, 0.2, ..., 0.9}.
decay factor = not reported
Distance decay in the RWMI update rule; optimized over {0.1, 0.2, ..., 0.9}.
k (latent dimension) = not reported
Rank of low-dimensional representations V, Q, B; not mentioned or specified in the paper.

assumptions (4)

domain assumption Mutual information between features and labels can be reliably estimated from finite samples.
Used to construct the MI matrix in Eq. (3); the estimation details (discretization, estimator choice) are omitted.
ad hoc to paper The random walk on the feature-label graph yields association weights that improve upon direct mutual information.
This is the central heuristic of the paper; no formal definition of indirect association or proof is provided.
domain assumption Data approximately lie in a shared low-dimensional latent space for features and labels.
Justifies the NMF formulation in Eq. (9) and the alignment term in Eq. (11).
domain assumption The multiplicative update rules converge to a useful optimum.
The paper claims global optimum for a non-convex objective, but only empirical convergence is shown, so the assumption is standard local convergence.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Graph Random Walk with Feature-Label Space Alignment: A Multi-Label Feature Selection Method." pith.science (2026). https://pith.science/paper/SE5LM3C6

@misc{pith2026250523228,
  author       = {Pith},
  title        = {Pith review of: Graph Random Walk with Feature-Label Space Alignment: A Multi-Label Feature Selection Method},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SE5LM3C6}},
  note         = {Machine review of arXiv:2505.23228}
}

read the original abstract

The rapid growth in feature dimension may introduce implicit associations between features and labels in multi-label datasets, making the relationships between features and labels increasingly complex. Moreover, existing methods often adopt low-dimensional linear decomposition to explore the associations between features and labels. However, linear decomposition struggles to capture complex nonlinear associations and may lead to misalignment between the feature space and the label space. To address these two critical challenges, we propose innovative solutions. First, we design a random walk graph that integrates feature-feature, label-label, and feature-label relationships to accurately capture nonlinear and implicit indirect associations, while optimizing the latent representations of associations between features and labels after low-rank decomposition. Second, we align the variable spaces by leveraging low-dimensional representation coefficients, while preserving the manifold structure between the original high-dimensional multi-label data and the low-dimensional representation space. Extensive experiments and ablation studies conducted on seven benchmark datasets and three representative datasets using various evaluation metrics demonstrate the superiority of the proposed method\footnote{Code: https://github.com/Heilong623/-GRW-}.

Figures

Figures reproduced from arXiv: 2505.23228 by the authors.

**Figure 2.** Algorithm framework. First, X and Y are decomposed using low-rank matrix factorization. Next, the feature-label association matrix is constrained by the random walk matrix Rw. Then, X and Y are aligned in a shared space. Finally, feature importance is computed via Q ⊤B. Feature-Label Relationship Update Rule Implicit indirect relationships refer to the potential high-order associations between feature nodes and la… view at source ↗

**Figure 3.** Eight methods on Arts. (a) α (b) β (c) γ (d) ϵ (e) δ [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Parameter sensitivity studies on the Arts dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Convergence curves on Arts and Business datasets. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 40 canonical work pages

[1]

Wrapper feature selection method based differential evolution and extreme learning machine for intrusion detection system

[Al-Yaseen et al., 2022] Wathiq Laftah Al-Yaseen, Ali Kad- hum Idrees, and Faezah Hamad Almasoudy. Wrapper feature selection method based differential evolution and extreme learning machine for intrusion detection system. Pattern Recognition, 132:108912,

work page 2022
[3]

Convex optimization

[Boyd and Vandenberghe, 2004] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge univer- sity press,

work page 2004
[10]

Multi-label feature selec- tion with shared common mode

[Hu et al., 2020] Liang Hu, Yonghao Li, Wanfu Gao, Ping Zhang, and Juncheng Hu. Multi-label feature selec- tion with shared common mode. Pattern Recognition, 104:107344,

work page 2020
[13]

Ex- ploiting multilabel information for noise-resilient feature selection

[Jian et al., 2018] Ling Jian, Jundong Li, and Huan Liu. Ex- ploiting multilabel information for noise-resilient feature selection. ACM Transactions on Intelligent Systems and Technology (TIST), 9(5):1–23,

work page 2018
[15]

Multi-label relieff and f-statistic feature selections for image annotation

[Kong et al., 2012] Deguang Kong, Chris Ding, Heng Huang, and Haifeng Zhao. Multi-label relieff and f-statistic feature selections for image annotation. In 2012 IEEE conference on computer vision and pattern recognition, pages 2352–2359. IEEE,

work page 2012
[17]

Mu- tual information-based multi-label feature selection using interaction information

[Lee and Kim, 2015] Jaesung Lee and Dae-Won Kim. Mu- tual information-based multi-label feature selection using interaction information. Expert Systems with Applications, 42(4):2013–2025,

work page 2015
[19]

Feature selection: A data perspective

[Li et al., 2017] Jundong Li, Kewei Cheng, Suhang Wang, Fred Morstatter, Robert P Trevino, Jiliang Tang, and Huan Liu. Feature selection: A data perspective. ACM comput- ing surveys (CSUR), 50(6):1–45,

work page 2017
[20]

Multi-label feature selection via robust flexible sparse reg- ularization

[Li et al., 2023] Yonghao Li, Liang Hu, and Wanfu Gao. Multi-label feature selection via robust flexible sparse reg- ularization. Pattern Recognition, 134:109074,

work page 2023

Show all 43 references

[21]

Multi-label feature selection with high-sparse personalized and low-redundancy shared common features.Information Processing & Management, 61(3):103633,

[Li et al., 2024] Yonghao Li, Liang Hu, and Wanfu Gao. Multi-label feature selection with high-sparse personalized and low-redundancy shared common features.Information Processing & Management, 61(3):103633,

2024
[23]

Multi-label feature selection based on max-dependency and min-redundancy

[Lin et al., 2015] Yaojin Lin, Qinghua Hu, Jinghua Liu, and Jie Duan. Multi-label feature selection based on max-dependency and min-redundancy. Neurocomputing, 168:92–103,

2015
[24]

Multi-view multi-label learning with high- order label correlation

[Liu et al., 2023] Bo Liu, Weibin Li, Yanshan Xiao, Xi- aodong Chen, Laiwang Liu, Changdong Liu, Kai Wang, and Peng Sun. Multi-view multi-label learning with high- order label correlation. Information Sciences , 624:165– 184,

2023
[25]

Efficient and robust feature selection via joint ℓ2, 1-norms minimization

[Nie et al., 2010] Feiping Nie, Heng Huang, Xiao Cai, and Chris Ding. Efficient and robust feature selection via joint ℓ2, 1-norms minimization. Advances in neural information processing systems, 23,

2010
[27]

A pruned problem transformation method for multi-label classification

[Read, 2008] Jesse Read. A pruned problem transformation method for multi-label classification. In Proc. 2008 New Zealand Computer Science Research Student Conference (NZCSRS 2008), volume 143150, page 41,

2008
[29]

A local random walk model for complex networks based on discriminative fea- ture combinations

[Song et al., 2019] Aibo Song, Yangyang Liu, Zhiang Wu, Mingyu Zhai, and Junzhou Luo. A local random walk model for complex networks based on discriminative fea- ture combinations. Expert Systems with Applications , 118:329–339,

2019
[31]

Unsupervised feature selection via latent representation learning and manifold regularization

[Tang et al., 2019] Chang Tang, Meiru Bian, Xinwang Liu, Miaomiao Li, Hua Zhou, Pichao Wang, and Hailin Yin. Unsupervised feature selection via latent representation learning and manifold regularization. Neural Networks, 117:163–178,

2019
[32]

Metafs: performance assessment of biomarker discovery in metaproteomics

[Tang et al., 2021] Jing Tang, Minjie Mou, Yunxia Wang, Yongchao Luo, and Feng Zhu. Metafs: performance assessment of biomarker discovery in metaproteomics. Briefings in Bioinformatics, 22(3):bbaa105,

2021
[33]

Mulan: A java library for multi-label learning.The Journal of Machine Learning Research, 12:2411–2414,

[Tsoumakas et al., 2011] Grigorios Tsoumakas, Eleftherios Spyromitros-Xioufis, Jozef Vilcek, and Ioannis Vlahavas. Mulan: A java library for multi-label learning.The Journal of Machine Learning Research, 12:2411–2414,

2011
[34]

Multi-label causal feature selection

[Wu et al., 2020] Xingyu Wu, Bingbing Jiang, Kui Yu, Huanhuan Chen, and Chunyan Miao. Multi-label causal feature selection. In Proceedings of the AAAI conference on artificial intelligence , volume 34, pages 6430–6437,

2020
[35]

Feature selection based on label distribution and fuzzy mutual information

[Xiong et al., 2021] Chuanzhen Xiong, Wenbin Qian, Ying- long Wang, and Jintao Huang. Feature selection based on label distribution and fuzzy mutual information. Informa- tion Sciences, 574:297–319,

2021
[36]

Feature selection using stochastic gates

[Yamada et al., 2020] Yutaro Yamada, Ofir Lindenbaum, Sa- hand Negahban, and Yuval Kluger. Feature selection using stochastic gates. In International conference on machine learning, pages 10648–10659. PMLR,

2020
[37]

Lle score: A new filter- based unsupervised feature selection method based on nonlinear manifold embedding and its application to im- age recognition

[Yao et al., 2017] Chao Yao, Ya-Feng Liu, Bo Jiang, Jun- gong Han, and Junwei Han. Lle score: A new filter- based unsupervised feature selection method based on nonlinear manifold embedding and its application to im- age recognition. IEEE Transactions on Image Processing, 26(1...

2017
[38]

Multilabel feature selection: a local causal structure learning approach

[Yu et al., 2021] Kui Yu, Mingzhu Cai, Xingyu Wu, Lin Liu, and Jiuyong Li. Multilabel feature selection: a local causal structure learning approach. IEEE Transactions on Neural Networks and Learning Systems, 34(6):3044–3057,

2021
[39]

Feature relevance term variation for multi-label feature selection

[Zhang and Gao, 2021] Ping Zhang and Wanfu Gao. Feature relevance term variation for multi-label feature selection. Applied Intelligence, 51:5095–5110,

2021
[40]

Distinguishing two types of labels for multi-label feature selection

[Zhang et al., 2019] Ping Zhang, Guixia Liu, and Wanfu Gao. Distinguishing two types of labels for multi-label feature selection. Pattern recognition, 95:72–82,

2019
[41]

A conditional-weight joint relevance metric for feature relevancy term

[Zhang et al., 2021] Ping Zhang, Wanfu Gao, Juncheng Hu, and Yonghao Li. A conditional-weight joint relevance metric for feature relevancy term. Engineering Applica- tions of Artificial Intelligence, 106:104481,

2021
[42]

Multi-label feature selection via latent representation learning and dynamic graph constraints

[Zhang et al., 2024] Yao Zhang, Wei Huo, and Jun Tang. Multi-label feature selection via latent representation learning and dynamic graph constraints. Pattern Recog- nition, 151:110411,

2024
[43]

Multi-label learning with emerging new labels

[Zhu et al., 2018] Yue Zhu, Kai Ming Ting, and Zhi-Hua Zhou. Multi-label learning with emerging new labels. IEEE Transactions on Knowledge and Data Engineering , 30(10):1901–1914, 2018

2018
[2004]

Exact top-k feature selection via l2, 0-norm constraint

[Cai et al., 2013] Xiao Cai, Feiping Nie, and Heng Huang. Exact top-k feature selection via l2, 0-norm constraint. In Twenty-third international joint conference on artificial in- telligence. Citeseer,

2013
[2008]

Feature selection in au- tomatic music genre classification

[Silla Jr et al., 2008] Carlos N Silla Jr, Alessandro L Ko- erich, and Celso AA Kaestner. Feature selection in au- tomatic music genre classification. In 2008 Tenth IEEE in- ternational symposium on multimedia, pages 39–44. IEEE,

2008
[2010]

Label enhancement-based fea- ture selection via fuzzy neighborhood discrimination in- dex

[Qian et al., 2022] Wenbin Qian, Chuanzhen Xiong, Yuhua Qian, and Yinglong Wang. Label enhancement-based fea- ture selection via fuzzy neighborhood discrimination in- dex. Knowledge-Based Systems, 250:109119,

2022
[2011]

Manifold learning with structured subspace for multi-label feature selection

[Fan et al., 2021] Yuling Fan, Jinghua Liu, Peizhong Liu, Yongzhao Du, Weiyao Lan, and Shunxiang Wu. Manifold learning with structured subspace for multi-label feature selection. Pattern Recognition, 120:108169,

2021
[2012]

A novel multi-label feature selection method with association rules and rough set.Information Sciences, 624:299–323,

[Kou et al., 2023] Yi Kou, Guoping Lin, Yuhua Qian, and Shujiao Liao. A novel multi-label feature selection method with association rules and rough set.Information Sciences, 624:299–323,

2023
[2013]

Few-sample feature selection via feature manifold learning

[Cohen et al., 2023] David Cohen, Tal Shnitzer, Yuval Kluger, and Ronen Talmon. Few-sample feature selection via feature manifold learning. In International Conference on Machine Learning, pages 6296–6319. PMLR,

2023
[2015]

Scls: Multi-label feature selection based on scalable criterion for large label set

[Lee and Kim, 2017] Jaesung Lee and Dae-Won Kim. Scls: Multi-label feature selection based on scalable criterion for large label set. Pattern Recognition, 66:342–352,

2017
[2016]

Multi-label informed feature selection

[Jian et al., 2016] Ling Jian, Jundong Li, Kai Shu, and Huan Liu. Multi-label informed feature selection. In IJCAI, vol- ume 16, pages 1627–33,

2016
[2017]

Multilabel feature selection with constrained latent struc- ture shared term

[Gao et al., 2023] Wanfu Gao, Yonghao Li, and Liang Hu. Multilabel feature selection with constrained latent struc- ture shared term. IEEE Transactions on Neural Networks and Learning Systems, 34(3):1253–1262,

2023
[2018]

Fea- ture selection based on absolute deviation factor for text classification

[Jin et al., 2023] Lingbin Jin, Li Zhang, and Lei Zhao. Fea- ture selection based on absolute deviation factor for text classification. Information Processing & Management , 60(3):103251,

2023
[2019]

Feature selection using fisher score and multilabel neighborhood rough sets for multilabel classification

[Sun et al., 2021] Lin Sun, Tianxiang Wang, Weiping Ding, Jiucheng Xu, and Yaojin Lin. Feature selection using fisher score and multilabel neighborhood rough sets for multilabel classification. Information Sciences, 578:887– 912,

2021
[2020]

Learning label-specific features and class-dependent labels for multi-label classification

[Huang et al., 2016] Jun Huang, Guorong Li, Qingming Huang, and Xindong Wu. Learning label-specific features and class-dependent labels for multi-label classification. IEEE transactions on knowledge and data engineering , 28(12):3309–3323,

2016
[2021]

Novel feature selection method based on random walk and artificial bee colony

[Feng et al., 2017] Lizhou Feng, Youwei Wang, and Wanli Zuo. Novel feature selection method based on random walk and artificial bee colony. Journal of Intelligent & Fuzzy Systems, 32(1):115–126,

2017
[2022]

Benchmark of filter methods for feature selection in high- dimensional gene expression survival data

[Bommert et al., 2022] Andrea Bommert, Thomas Wel- chowski, Matthias Schmid, and J ¨org Rahnenf ¨uhrer. Benchmark of filter methods for feature selection in high- dimensional gene expression survival data. Briefings in Bioinformatics, 23(1):bbab354,

2022
[2023]

Feature selection for multi-label classification problems

[Doquire and Verleysen, 2011] Gauthier Doquire and Michel Verleysen. Feature selection for multi-label classification problems. In Advances in Computational Intelligence: 11th International Work-Conference on Artificial Neural Networks, IWANN 2011, Torremolinos- M´alaga, Spain...

2011
[2024]

Mfc: Initialization method for multi-label feature selection based on conditional mutual information

[Lim and Kim, 2020] Hyunki Lim and Dae-Won Kim. Mfc: Initialization method for multi-label feature selection based on conditional mutual information. Neurocomput- ing, 382:40–51,

2020

Pith tools

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

[1] [1]

Wrapper feature selection method based differential evolution and extreme learning machine for intrusion detection system

[Al-Yaseen et al., 2022] Wathiq Laftah Al-Yaseen, Ali Kad- hum Idrees, and Faezah Hamad Almasoudy. Wrapper feature selection method based differential evolution and extreme learning machine for intrusion detection system. Pattern Recognition, 132:108912,

work page 2022

[2] [3]

Convex optimization

[Boyd and Vandenberghe, 2004] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge univer- sity press,

work page 2004

[3] [10]

Multi-label feature selec- tion with shared common mode

[Hu et al., 2020] Liang Hu, Yonghao Li, Wanfu Gao, Ping Zhang, and Juncheng Hu. Multi-label feature selec- tion with shared common mode. Pattern Recognition, 104:107344,

work page 2020

[4] [13]

Ex- ploiting multilabel information for noise-resilient feature selection

[Jian et al., 2018] Ling Jian, Jundong Li, and Huan Liu. Ex- ploiting multilabel information for noise-resilient feature selection. ACM Transactions on Intelligent Systems and Technology (TIST), 9(5):1–23,

work page 2018

[5] [15]

Multi-label relieff and f-statistic feature selections for image annotation

[Kong et al., 2012] Deguang Kong, Chris Ding, Heng Huang, and Haifeng Zhao. Multi-label relieff and f-statistic feature selections for image annotation. In 2012 IEEE conference on computer vision and pattern recognition, pages 2352–2359. IEEE,

work page 2012

[6] [17]

Mu- tual information-based multi-label feature selection using interaction information

[Lee and Kim, 2015] Jaesung Lee and Dae-Won Kim. Mu- tual information-based multi-label feature selection using interaction information. Expert Systems with Applications, 42(4):2013–2025,

work page 2015

[7] [19]

Feature selection: A data perspective

[Li et al., 2017] Jundong Li, Kewei Cheng, Suhang Wang, Fred Morstatter, Robert P Trevino, Jiliang Tang, and Huan Liu. Feature selection: A data perspective. ACM comput- ing surveys (CSUR), 50(6):1–45,

work page 2017

[8] [20]

Multi-label feature selection via robust flexible sparse reg- ularization

[Li et al., 2023] Yonghao Li, Liang Hu, and Wanfu Gao. Multi-label feature selection via robust flexible sparse reg- ularization. Pattern Recognition, 134:109074,

work page 2023