Pith. sign in

REVIEW 4 major objections 6 minor 56 references

Constrained Non-negative Matrix Factorization for Guided Topic Modeling of Minority Topics

T0 review · 4 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 single seed list plus two constraints recovers rare topics other models miss.

desk verdict A plausible new seeded-NMF variant for minority-topic discovery, but the W constraint quietly assumes minority content only appears in documents containing exact seed words, and the synthetic evaluation probably never tests that assumption. read the letter →

arxiv 2505.16493 v1 pith:E7WJYSSO submitted 2025-05-22 cs.LG

Seyedeh Fatemeh Ebrahimi , Jaakko Peltonen This is my paper
classification cs.LG
keywords topicmodelingnon-negativematrixfactorizationminoritytopicsseedwordsconstrainedoptimizationKKTconditionsmentalhealthYouTubecomments
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 tries to establish that a single hand-supplied seed word list plus two mild inequality constraints is enough to make non-negative matrix factorization recover rare “minority” topics that standard topic models drown out. The motivation is domain-critical content such as mental-health discussion, which appears in only a small share of YouTube comments and often only briefly within a comment. The constraints are soft: documents containing none of the seed words are capped at nearly zero prevalence of each minority topic, and each minority topic must devote at least a user-set share of its word mass to seed words. On a synthetic corpus with injected ground-truth minority topics, the method reports the highest purity (0.1765) and normalized mutual information (0.2453) among eleven baselines, and the smallest Jensen-Shannon divergence from the true topics. A case study on Finnish YouTube vlog comments shows the discovered minority topics are interpretable mental-health themes such as suicide, ADHD support, and holiday-related depression.

What carries the argument

The load-bearing object is the pair of inequality constraints on the two NMF factors. The constraint $g_{1,ik}(W)=W_{ik}-W_{\max}\le 0$ applies only to documents that contain no seed words and to minority-topic columns, so vocabulary outside the seed list cannot push minority prevalence upward. The constraint $g_{2,k}(H)=\theta_{\min}-\frac{\sum_{j'\in S_I} H_{kj'}}{\sum_{j'=1}^{N} H_{kj'}}\le 0$ anchors each minority topic to the seed content collectively, without dictating which seed words belong to which topic. Optimization minimizes the generalized KL divergence subject to these constraints through a Lagrangian; setting the Lagrangian gradients to zero gives multiplicative updates for $W$, $H$ and for the Lagrange multipliers $\lambda$ and $\mu$, each update being a ratio of the positive and negative parts of the gradient, which preserves non-negativity.

What would settle it

Build a synthetic corpus whose injected minority topics use synonyms and paraphrases that are deliberately absent from the supplied seed list, with ground-truth labels known. If the constrained NMF at $W_{\max}=10^{-9}$ assigns those documents to minority topics no better than an unconstrained NMF does (purity and NMI near baseline), the coverage assumption behind the W constraint fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that a constrained NMF can separate data-driven minority topics from majority topics without topic-level supervision. The expert supplies one seed list and two scalars: a cap $W_{\max}$ on minority-topic prevalence in documents that contain no seed words, and a floor $\theta_{\min}$ on the total share of seed words in each minority topic's word distribution. From a Lagrangian of the generalized KL divergence under these constraints, Karush-Kuhn-Tucker stationarity yields multiplicative update rules for $W$, $H$ and the Lagrange multipliers $\lambda$ and $\mu$; the authors argue these updates guarantee the constraints are satisfied. Empirically, the method gives the best clustering and closest topic-word distributions to ground truth on the synthetic benchmark, and it produces coherent mental-health topics in real YouTube comments. The authors also note that convergence theorems for the constrained updates are left to future work and that automated topic-quality metrics can be misaligned with semantic value.

Load-bearing premise

The load-bearing assumption is that a document containing none of the seed words has essentially no minority content: with $W_{\max}$ set to $10^{-9}$, such documents are forced to near-zero prevalence in every minority topic, so any minority theme expressed in vocabulary outside the seed list is excluded by construction.

Editorial extensions

If this is right

  • Domain experts can guide discovery of rare content with a single seed list, without pre-specifying how that content divides into subtopics.
  • Low-prevalence themes can be modeled even when the seed words themselves are not prominent within the minority topics, because the constraints are only on aggregate seed-word share.
  • On the synthetic benchmark the method beats both unsupervised and seed-guided baselines on purity, NMI, and Jensen-Shannon divergence, and its relative edge grows as the number of total and minority topics increases.
  • The constrained formulation works on any non-negative document-term matrix, so the same pair of constraints transfers to other imbalanced corpora beyond the mental-health case study.

Reading between the lines

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

  • A diagnostic reading of the W constraint: documents without seed words that still receive minority prevalence after fitting are exactly the cases where the seed list is incomplete; reporting them could serve as an active-learning signal for vocabulary expansion.
  • If the seed list has imperfect coverage, the fixed $W_{\max}=10^{-9}$ will suppress minority content expressed through paraphrase or slang; a testable extension is to set $W_{\max}$ from a small validation set or to relax it for documents whose embeddings are close to seed words.
  • The same KKT-style derivation should transfer to other constrained matrix-factorization settings where one factor needs elementwise upper bounds and another needs row-wise concentration constraints, such as source separation with labeled anchors.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a constrained non-negative matrix factorization (NMF) method for guided topic modeling of minority topics. The user provides a single seed word list for the minority domain (e.g., mental health). Two inequality constraints are introduced: g1 (Eq. 1) caps the prevalence of minority topics in documents that contain no seed words, and g2 (Eq. 2) requires each minority topic to devote at least a fraction theta_min of its word mass to seed words. The model is fitted by minimizing a generalized KL divergence under these constraints, with multiplicative updates for W and H derived from KKT conditions and gradient-ascent updates for Lagrange multipliers. Experiments on a synthetic dataset report higher NMI, purity, and lower JSD than several baselines, and a case study on Finnish YouTube comments yields interpretable mental-health topics. The paper also provides an appendix with detailed derivations, ablations, and additional results.

Significance. If the method works as claimed, it offers a practical way to inject domain knowledge through a single seed list without topic-level supervision or predefined topic divisions. The idea of separating minority and majority topics with soft prevalence constraints is sensible and could be useful for analyzing low-prevalence themes in large text collections. The paper includes a thorough derivation appendix and compares against a broad set of baselines, which are strengths. However, the empirical evaluation is weakened by a likely overlap between the synthetic injected vocabulary and the seed list, and the theoretical claims about guaranteed constraint satisfaction and nonnegativity are stronger than what is actually proven. The method's simplicity and interpretability, if the issues are addressed, would make it a solid contribution to guided topic modeling.

major comments (4)
  1. [Section 4] The statement "Our updates derived using KKT conditions guarantee constraints are satisfied" is not supported. The KKT conditions are used only to derive stationarity conditions; the constraints are enforced via Lagrange multipliers updated by gradient ascent (Eqs. 12-13), with no convergence proof. The paper itself concedes that convergence is unproven and that constraints are not satisfied at every iteration. At best, the update rules are KKT-inspired and empirically satisfy constraints at convergence. This overclaim should be removed or explicitly qualified.
  2. [Appendix A.8, Eq. (21) and Algorithm 2] The update rule for H_kj in Eq. (21) has a denominator containing mu_k * (Num_k/Den_k^2 - delta_{j in S_I}/Den_k), which can be negative for seed columns j when mu_k is large. Consequently, nonnegativity of H is not guaranteed, contradicting the claim in Section 4 that the updates retain nonnegativity of W and H. The alternate rule in Eq. (22) does guarantee nonnegativity, but it is not the rule used in the experiments or in Algorithm 2. The paper should either adopt Eq. (22) as the default or provide a bound on mu_k that keeps the denominator in Eq. (21) positive.
  3. [Section 5.1 and Appendix A.11, Eq. (1)] The constraint Wmax = 1e-9 encodes the assumption that documents without any seed word have essentially zero prevalence in minority topics. The synthetic data is generated by injecting words from a predefined 18-topic vocabulary, but the paper does not report how much this injected vocabulary overlaps with the user-provided seed list. If the overlap is high, then almost all minority documents contain seed words, I0 contains few minority documents, and the constraint's harmful effect is never exercised. The paper's claim that the seed list need not be comprehensive is thus undermined by the extreme Wmax value. The authors should report the overlap rate and add a synthetic condition where minority content is expressed through non-seed or paraphrased vocabulary.
  4. [Section 5.2, focused purity] The customized purity metric divides by sum_k delta(count(y_k, k) > 0) * count(k), which excludes clusters that contain no minority documents from the denominator. As written, a method that places minority documents into pure clusters and splits the rest arbitrarily can achieve a high score without meaningful separation of minority documents. This makes the reported purity improvements hard to interpret. Please also report standard purity (computed over all documents) or provide a rationale for excluding background-only clusters.
minor comments (6)
  1. [Section 1] The roadmap sentence says "Section 5.3 gives results and A.12 findings"; the appendix reference should be to "Section A.12".
  2. [Eq. (3) and Section 3.2] The Lagrangian is written as lambda * g1(W) + mu * g2(H), but the subsequent derivation treats g1 and g2 as sums of per-element constraints with individual multipliers. Please make the notation consistent, e.g., by writing explicit sums over i,k from the start.
  3. [Tables 2-5] The description says the real-dataset topics are split across four tables, but it is not stated how the 50 topics are partitioned across these tables; adding the topic ranges in each caption would improve clarity.
  4. [Appendix A.12] The text refers to "Section 6 of the main paper" when discussing results; the relevant section is Section 5.3.
  5. [Appendix A.13] Error bars are reported only for KL divergence across the 10 runs, not for NMI or purity. Since the main claims are about NMI/purity, reporting their mean and standard deviation would strengthen the robustness argument.
  6. [Appendix A.11] The seed word list is said to be available in a GitHub repository, but the repository is not named in the manuscript text; please provide a non-anonymous link or include the seed list in the supplement so that the experiments are reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No demonstrated circularity: the constraints are explicit modeling assumptions and the synthetic benchmark overlap is not shown, so no load-bearing step reduces to its own inputs.

full rationale

I walked the paper's derivation chain. The method minimizes generalized KL divergence under two inequality constraints: g1 caps minority-topic prevalence in documents containing no seed words (Eq. 1), and g2 requires a minimum total seed-word share in each minority topic (Eq. 2). These are stated modeling assumptions, not quantities fitted from the data and then relabeled as predictions. The multiplicative updates are derived from KKT stationarity conditions and the constraint gradients; the update rules follow algebraically from the Lagrangian and are not circular. The synthetic evaluation is generated by injecting words from predefined mental-health topics, and the model is evaluated against those injected labels. A possible circularity would arise if the injected vocabulary were identical to the seed list used in Eq. 1, because then documents labeled as minority are exactly those permitted to carry minority prevalence. However, the paper does not state that the injected words are the seed words; it describes the synthetic topics as having 'related Finnish words' and the seed list as a separate collection of mental-health terms. Without a quoted or equation-level reduction showing that the ground-truth minority labels are equivalent to seed-list membership by construction, the concern is an experimental-validity risk rather than an exhibited circularity. The paper also explicitly acknowledges limitations: automatic coherence-style measures 'risk being misaligned or biased,' and convergence theorems are 'left to future work.' Those are honest caveats, not circular reasoning. I therefore find no load-bearing circular step and score the paper 0.

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

The central claim rests on two user-set constraints (Wmax, theta_min), the unstated overlap between the seed list and the synthetic injected vocabulary, and the unproven convergence of the KKT-based updates. No new physical or latent entities are introduced; the method is a constrained factorization, so the ledger consists of modeling assumptions and fitted hyperparameters rather than new ontology.

free parameters (4)
  • Wmax = 1e-9
    Upper bound on prevalence of each minority topic in documents without seed words; tuned to be 'optimal across trials' on the synthetic dataset (A.11), which is also the evaluation set.
  • theta_min = 0.4
    Lower bound on the proportion of seed words in each minority topic's word distribution; user-set and used in all experiments, forcing seed words to be prominent (Eq. 2).
  • eta (learning rate) = 0.001
    Learning rate for Lagrange multiplier updates; chosen by hand (A.11).
  • K_MH = one-third of K
    Number of minority topics, set as one-third of total topics in experiments, a modeling choice that determines granularity.
assumptions (5)
  • standard math Standard NMF with generalized KL divergence is an appropriate generative model for document-term counts.
    Invoked in Section 3; it is a standard factorization model.
  • ad hoc to paper Documents without any seed word should have near-zero prevalence of minority topics.
    This is the g1 constraint, Eq. (1), with Wmax=1e-9 in A.11; it is a domain assumption imposed by the authors, not derived from data.
  • ad hoc to paper Each minority topic must concentrate at least theta_min=0.4 of its word mass on the seed list.
    This is the g2 constraint, Eq. (2), with theta_min=0.4 in A.11; it contradicts the paper's claim that seed words need not be prominent.
  • standard math KKT conditions describe the optimum of the constrained NMF problem and the proposed multiplicative updates approach it.
    Used in Section 3.3; the KKT framework is standard, but convergence of the updates is acknowledged as unproven.
  • domain assumption The injected words in the synthetic corpus are covered by the seed word list.
    Unstated but load-bearing for the synthetic evaluation; the paper gives examples like 'itsemurha' and 'ahdistus' that are typical mental-health seed words, yet never measures overlap between the synthetic ground-truth words and the seed list (Sections 5.1, A.9).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Constrained Non-negative Matrix Factorization for Guided Topic Modeling of Minority Topics." pith.science (2026). https://pith.science/paper/E7WJYSSO

@misc{pith2026250516493,
  author       = {Pith},
  title        = {Pith review of: Constrained Non-negative Matrix Factorization for Guided Topic Modeling of Minority Topics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E7WJYSSO}},
  note         = {Machine review of arXiv:2505.16493}
}
read the original abstract

Topic models often fail to capture low-prevalence, domain-critical themes, so-called minority topics, such as mental health themes in online comments. While some existing methods can incorporate domain knowledge, such as expected topical content, methods allowing guidance may require overly detailed expected topics, hindering the discovery of topic divisions and variation. We propose a topic modeling solution via a specially constrained NMF. We incorporate a seed word list characterizing minority content of interest, but we do not require experts to pre-specify their division across minority topics. Through prevalence constraints on minority topics and seed word content across topics, we learn distinct data-driven minority topics as well as majority topics. The constrained NMF is fitted via Karush-Kuhn-Tucker (KKT) conditions with multiplicative updates. We outperform several baselines on synthetic data in terms of topic purity, normalized mutual information, and also evaluate topic quality using Jensen-Shannon divergence (JSD). We conduct a case study on YouTube vlog comments, analyzing viewer discussion of mental health content; our model successfully identifies and reveals this domain-relevant minority content.

Figures

Figures reproduced from arXiv: 2505.16493 by the authors.

Figure 2
Figure 2. reports the topic quality of the mod￾els according to JSD. Here as well, our method yielded best performance (smallest JSDs) in dis￾covering minority themes. Thus we achieve both high topic quality (small JSD) and high clustering ability (high purity & NMI) outperforming others. The detailed topics discovered by our model are shown in Appendix A.12, [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 1
Figure 1. Comparison of NMI and Purity Scores across [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 3
Figure 3. KL Divergence Across Iterations with Error [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Topic Quality using JSD Score [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]
Figure 5
Figure 5. Figure 5: Effect of Wmax and θmin on NMI and purity scores [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]
Figure 6
Figure 6. Figure 6: Effect of Wmax and θmin on KL divergence [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]
Figure 7
Figure 7. Figure 7: NMI scores for topic counts 30, 50, and 80. [PITH_FULL_IMAGE:figures/full_fig_p024_7.png]
Figure 8
Figure 8. Figure 8: Purity scores for topic counts 30, 50, and 80. [PITH_FULL_IMAGE:figures/full_fig_p024_8.png]
Figure 9
Figure 9. Figure 9: Scatter plot of Purity vs. NMI scores for different models. [PITH_FULL_IMAGE:figures/full_fig_p025_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 39 canonical work pages

  1. [1]

    Rania Albalawi, Tet Hin Yeap, and Morad Benyoucef. 2020. https://api.semanticscholar.org/CorpusID:220497670 Using topic modeling methods for short-text data: A comparative analysis . Frontiers in Artificial Intelligence, 3

    work page 2020

  2. [2]

    Dimitar Angelov. 2020. https://api.semanticscholar.org/CorpusID:221246303 Top2vec: Distributed representations of topics . ArXiv, abs/2008.09470

    arXiv 2020

  3. [3]

    Berry, Murray Browne, Amy N

    Michael W. Berry, Murray Browne, Amy N. Langville, V. Paul Pauca, and Robert J. Plemmons. 2007. https://doi.org/10.1016/j.csda.2006.11.006 Algorithms and applications for approximate nonnegative matrix factorization . Computational Statistics & Data Analysis, 52(1):155--173

    doi:10.1016/j.csda.2006.11.006 2007

  4. [4]

    Federico Bianchi, Silvia Terragni, and Dirk Hovy. 2021 a . https://doi.org/10.18653/v1/2021.acl-short.96 Pre-training is a hot topic: Contextualized document embeddings improve topic coherence . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Process...

    doi:10.18653/v1/2021.acl-short.96 2021

  5. [5]

    Federico Bianchi, Silvia Terragni, Dirk Hovy, Debora Nozza, and Elisabetta Fersini. 2021 b . https://doi.org/10.18653/v1/2021.eacl-main.143 Cross-lingual contextualized topic models with zero-shot learning . In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1676--1683, Online...

    doi:10.18653/v1/2021.eacl-main.143 2021

  6. [6]

    Blei, Andrew Y

    David M. Blei, Andrew Y. Ng, and Michael I. Jordan. 2003. Latent D irichlet allocation. J. Mach. Learn. Res., 3(null):993–1022

    work page 2003

  7. [7]

    Peter Carbonetto, Abhishek Sarkar, Zihao Wang, and Matthew Stephens. 2021. Non-negative matrix factorization algorithms greatly improve topic model fits. arXiv preprint arXiv:2105.13440

    work page arXiv 2021

  8. [8]

    Dallas Card, Chenhao Tan, and Noah A. Smith. 2018. https://doi.org/10.18653/v1/P18-1189 Neural models for documents with metadata . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2031--2040, Melbourne, Australia. Association for Computational Linguistics

    work page doi:10.18653/v1/p18-1189 2018

Show all 56 references
  1. [9]

    Jonathan Chang, Sean Gerrish, Chong Wang, Jordan Boyd-graber, and David Blei. 2009. https://proceedings.neurips.cc/paper_files/paper/2009/file/f92586a25bb3145facd64ab20fd554ff-Paper.pdf Reading tea leaves: How humans interpret topic models . In Advances in Neural Information P...

    2009

  2. [10]

    Yong Chen, Hui Zhang, Rui Liu, Zhiwen Ye, and Jianying Lin. 2019. Experimental explorations on short text topic mining between LDA and NMF based schemes. Knowledge-Based Systems, 163:1--13

    2019

  3. [11]

    D Manning Christopher, Raghavan Prabhakar, and Schutze Hinrich. 2008. Introduction to information retrieval

    2008

  4. [12]

    Mrinal Das, Suparna Bhattacharya, Chiranjib Bhattacharyya, and Gopinath Kanchi. 2013. https://proceedings.mlr.press/v28/das13.html Subtle topic models and discovering subtly manifested software concerns automatically . In Proceedings of the 30th International Conference on Mac...

    2013

  5. [13]

    Mrinal Das and Gaurav Jain. 2024. Human guided multi-proportions topic model for rare event detection without using labels. In Intelligent Systems and Applications, pages 368--385, Cham. Springer Nature Switzerland

    2024

  6. [14]

    Roman Egger and Joanne Yu. 2022. https://api.semanticscholar.org/CorpusID:248530058 A topic modeling comparison between LDA , NMF , T op2 V ec, and BERT opic to demystify T witter posts . Frontiers in Sociology, 7

    2022

  7. [15]

    Jacob Eisenstein, Amr Ahmed, and Eric P Xing. 2011. Sparse additive generative models of text. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 1041--1048

    2011

  8. [16]

    Shusei Eshima, Kosuke Imai, and Tomoya Sasaki. 2023. https://doi.org/10.1111/ajps.12779 Keyword‐assisted topic models . American Journal of Political Science, 68

    2023 doi

  9. [17]

    Lorenzo Finesso and Peter Spreij. 2006. https://doi.org/10.1016/j.laa.2005.11.012 Nonnegative matrix factorization and I -divergence alternating minimization . Linear Algebra and its Applications, 416(2–3):270–287

    2006 doi

  10. [18]

    Gallagher, Kyle Reing, David C

    Ryan J. Gallagher, Kyle Reing, David C. Kale, and Greg Ver Steeg. 2016. https://api.semanticscholar.org/CorpusID:18449288 Anchored correlation explanation: Topic modeling with minimal domain knowledge . Transactions of the Association for Computational Linguistics, 5:529--542

    2016

  11. [19]

    Benyamin Ghojogh, Ali Ghodsi, Fakhri Karray, and Mark Crowley. 2021. KKT conditions, first-order and second-order optimization, and distributed optimization: tutorial and survey. arXiv preprint arXiv:2110.01858

    2021 arXiv

  12. [20]

    Edward F Gonzalez and Yin Zhang. 2005. Accelerating the lee-seung algorithm for non-negative matrix factorization. Dept. Comput. & Appl. Math., Rice Univ., Houston, TX, Tech. Rep. TR-05-02, pages 1--13

    2005

  13. [21]

    Maarten Grootendorst. 2022. Bertopic: Neural topic modeling with a class-based tf-idf procedure. arXiv preprint arXiv:2203.05794

    2022 arXiv

  14. [22]

    Jamie Haddock, Lara Kassab, Sixian Li, Alona Kryshchenko, Rachel Grotheer, Elena Sizikova, Chuntian Wang, Thomas Merkh, R. W. M. A. Madushani, Miju Ahn, Deanna Needell, and Kathryn Leonard. 2020. https://api.semanticscholar.org/CorpusID:223953533 Semi-supervised NMF models for...

    2020 arXiv

  15. [23]

    Thomas Hofmann. 1999. https://doi.org/10.1145/312624.312649 Probabilistic latent semantic indexing . In Proceedings of the 22nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '99, page 50–57, New York, NY, USA. Association...

    1999

  16. [24]

    Alexander Hoyle, Pranav Goel, Andrew Hian-Cheong, Denis Peskov, Jordan Boyd-Graber, and Philip Resnik. 2021. Is automated topic model evaluation broken? the incoherence of coherence. Advances in neural information processing systems, 34:2018--2033

    2021

  17. [25]

    Jagadeesh Jagarlamudi, Hal Daum \'e III, and Raghavendra Udupa. 2012. Incorporating lexical priors into topic models. In Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 204--213

    2012

  18. [26]

    Yuheng Jia, Sam Kwong, Junhui Hou, and Wenhui Wu. 2020. https://doi.org/10.1109/TNNLS.2019.2933223 Semi-supervised non-negative matrix factorization with dissimilarity and similarity regularization . IEEE Transactions on Neural Networks and Learning Systems, 31(7):2510--2521

    2020

  19. [27]

    Yuheng Jia, Hui Liu, Junhui Hou, and Sam Kwong. 2021. https://doi.org/10.1109/TCYB.2020.2969684 Semisupervised adaptive symmetric non-negative matrix factorization . IEEE Transactions on Cybernetics, 51(5):2550--2562

    2021

  20. [28]

    James M. Joyce. 2011. https://api.semanticscholar.org/CorpusID:37718089 Kullback- L eibler divergence . In International Encyclopedia of Statistical Science

    2011

  21. [29]

    Kenneth Lange. 2013. https://doi.org/10.1007/978-1-4614-5838-8_5 Karush-Kuhn-Tucker Theory , pages 107--135. Springer New York, New York, NY

    2013 doi

  22. [30]

    Daniel Lee and H Sebastian Seung. 2000. Algorithms for non-negative matrix factorization. Advances in neural information processing systems, 13

    2000

  23. [31]

    Dongha Lee, Jiaming Shen, SeongKu Kang, Susik Yoon, Jiawei Han, and Hwanjo Yu. 2022 a . Taxocom: Topic taxonomy completion with hierarchical discovery of novel topic clusters. In Proceedings of the ACM Web Conference 2022, pages 2819--2829

    2022

  24. [32]

    Dongha Lee, Jiaming Shen, Seonghyeon Lee, Susik Yoon, Hwanjo Yu, and Jiawei Han. 2022 b . Topic taxonomy expansion via hierarchy-aware topic phrase generation. arXiv preprint arXiv:2211.01981

    2022 arXiv

  25. [33]

    Hyekyoung Lee, Jiho Yoo, and Seungjin Choi. 2010. https://doi.org/10.1109/LSP.2009.2027163 Semi-supervised nonnegative matrix factorization . IEEE Signal Processing Letters, 17:4--7

    2010

  26. [34]

    J. W. Leech. 1965. https://doi.org/10.1007/978-94-010-9169-5_3 The Lagrangian Formulation , pages 17--25. Springer Netherlands, Dordrecht

    1965 doi

  27. [35]

    Chew, Longxiu Huang, Benjamin Jarman, and Deanna Needell

    Pengyu Li, Christine Tseng, Yaxuan Zheng, Joyce A. Chew, Longxiu Huang, Benjamin Jarman, and Deanna Needell. 2022. https://doi.org/10.3390/a15050136 Guided semi-supervised non-negative matrix factorization . Algorithms, 15(5)

    2022 doi

  28. [36]

    Chih-Jen Lin. 2007. https://doi.org/10.1109/TNN.2007.895831 On the convergence of multiplicative update algorithms for nonnegative matrix factorization . IEEE Transactions on Neural Networks, 18(6):1589--1596

    2007

  29. [37]

    Yang Lin, Xin Gao, Xu Chu, Yasha Wang, Junfeng Zhao, and Chao Chen. 2023. Enhancing neural topic model with multi-level supervisions from seed words. In Findings of the Association for Computational Linguistics: ACL 2023, pages 13361--13377

    2023

  30. [38]

    Lindstrom, Xiaofu Ding, Feng Liu, Anand Somayajula, and Deanna Needell

    Michael R. Lindstrom, Xiaofu Ding, Feng Liu, Anand Somayajula, and Deanna Needell. 2022. https://api.semanticscholar.org/CorpusID:254877130 Continuous semi-supervised nonnegative matrix factorization . Algorithms, 16:187

    2022

  31. [39]

    Jiyuan Liu, Hegang Chen, Chunjiang Zhu, and Yanghui Rao. 2024. Unsupervised hierarchical topic modeling via anchor word clustering and path guidance. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 7505--7517

    2024

  32. [40]

    Adewale Obadimu, Esther Mead, and Nitin Agarwal. 2019. Identifying latent toxic features on Y ou T ube using non-negative matrix factorization. In The Ninth International Conference on Social Media Technologies, Communication, and Informatics, IEEE

    2019

  33. [41]

    Duy-Tung Pham, Thien Trang Nguyen Vu, Tung Nguyen, Linh Van Ngo, Duc Anh Nguyen, and Thien Huu Nguyen. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.457 N euro M ax: Enhancing neural topic modeling via maximizing mutual information and group topic regularization . In F...

    2024 doi

  34. [42]

    Gerard Salton and Christopher Buckley. 1988. https://doi.org/10.1016/0306-4573(88)90021-0 Term-weighting approaches in automatic text retrieval . Information Processing & Management, 24(5):513--523

    1988 doi

  35. [43]

    Suzanna Sia, Ayush Dalmia, and Sabrina J. Mielke. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.135 Tired of topic models? clusters of pretrained word embeddings make for fast and good topics too! In Proceedings of the 2020 Conference on Empirical Methods in Natural Langua...

    2020 doi

  36. [44]

    Akash Srivastava and Charles Sutton. 2017. https://api.semanticscholar.org/CorpusID:29842525 Autoencoding variational inference for topic models . In International Conference on Learning Representations

    2017

  37. [45]

    Greg Ver Steeg and A. G. Galstyan. 2014. https://api.semanticscholar.org/CorpusID:14240005 Discovering structure in high-dimensional data through correlation explanation . In Neural Information Processing Systems

    2014

  38. [46]

    Joshua Vendrow, Jamie Haddock, Elizaveta Rebrova, and Deanna Needell. 2021. https://doi.org/10.1109/ICASSP39728.2021.9413656 On a guided nonnegative matrix factorization . In ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), ...

    2021

  39. [47]

    Yu-Xiong Wang and Yu-Jin Zhang. 2013. https://doi.org/10.1109/TKDE.2012.51 Nonnegative matrix factorization: A comprehensive review . IEEE Transactions on Knowledge and Data Engineering, 25(6):1336--1353

    2013 doi

  40. [48]

    Xiaobao Wu, Xinshuai Dong, Thong Thanh Nguyen, and Anh Tuan Luu. 2023. Effective neural topic modeling with embedding clustering regularization. In International Conference on Machine Learning, pages 37335--37357. PMLR

    2023

  41. [49]

    Xiaobao Wu, Thong Nguyen, and Anh Tuan Luu. 2024 a . https://api.semanticscholar.org/CorpusID:267297321 A survey on neural topic models: Methods, applications, and challenges . ArXiv, abs/2401.15351

    2024 arXiv

  42. [50]

    Xiaobao Wu, Thong Thanh Nguyen, Delvin Ce Zhang, William Yang Wang, and Anh Tuan Luu. 2024 b . https://openreview.net/forum?id=7t6aq0Fa9D FAST opic: Pretrained transformer is a fast, adaptive, stable, and transferable topic model . In The Thirty-eighth Annual Conference on Neu...

    2024

  43. [51]

    Yu Zhang, Yunyi Zhang, Martin Michalski, Yucheng Jiang, Yu Meng, and Jiawei Han. 2023. https://doi.org/10.1145/3539597.3570475 Effective seed-guided topic discovery by integrating multiple types of contexts . In Proceedings of the Sixteenth ACM International Conference on Web ...

    2023

  44. [52]

    Zhong-Yuan Zhang. 2012. Nonnegative matrix factorization: models, algorithms and applications. Data Mining: Foundations and Intelligent Paradigms: Volume 2: Statistical, Bayesian, Time Series and other Theoretical Aspects, pages 99--134

    2012

  45. [53]

    He Zhao, Dinh Phung, Viet Huynh, Yuan Jin, Lan Du, and Wray Buntine. 2021 a . Topic modelling meets deep neural networks: A survey. arXiv preprint arXiv:2103.00498

    2021 arXiv

  46. [54]

    He Zhao, Dinh Phung, Viet Huynh, Trung Le, and Wray Buntine. 2021 b . https://openreview.net/forum?id=Oos98K9Lv-k Neural topic model via optimal transport . In International Conference on Learning Representations

    2021

  47. [55]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  48. [56]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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