Topic Modeling and Link-Prediction for Material Property Discovery

Boian S. Alexandrov; Cynthia Matuszek; Maksim E. Eren; Ryan C. Barron; Valentin Stanev

REVIEW 4 major objections 6 minor 18 references

Topic Modeling and Link-Prediction for Material Property Discovery

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that masking every known superconductivity link in a topic–material matrix still lets a BNMFk+LMF ensemble rank all four benchmark superconducting TMDs at the top.

desk verdict A useful application of existing matrix factorization tools to literature-based materials discovery, but the headline validation is internal consistency, not external prediction. read the letter →

arxiv 2507.06139 v1 pith:SP2F4BW7 submitted 2025-07-08 cs.LG cs.AIcs.CE

Ryan C. Barron , Maksim E. Eren , Valentin Stanev , Cynthia Matuszek , Boian S. Alexandrov This is my paper · ORCID

classification cs.LGcs.AIcs.CE

keywords hierarchicalnonnegativematrixfactorizationbooleanlogisticlinkpredictionmaterialpropertydiscoverytransition-metaldichalcogenidesscientificliteratureminingcompletion

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 aims to show that hidden material-property relationships can be recovered from the latent topic structure of a scientific corpus alone. Working with 46,862 papers on transition-metal dichalcogenides, the authors build a three-level hierarchy of research topics, connect those topics to 72 materials in a binary Materials Property Matrix, and then mask every known superconductivity link for four benchmark compounds. Their BNMFk-LMF ensemble recovers all four masked superconductors within the top three predictions and ranks them above all non-superconductors, with positive scores concentrated near 1 and negatives near 0. If this holds, the pipeline gives materials scientists a literature-driven way to rank candidate compounds for experimental study before synthesis, which matters because only 72 of a vast combinatorial space of TMDs are known.

What carries the argument

The load-bearing object is the Materials Property Matrix, a 815×72 binary matrix whose entry is 1 when a material is associated with a latent topic discovered by HNMFk, 0 when no association is observed, and missing when the literature is inconclusive. On top of that matrix, the BNMFk-LMF ensemble combines a Boolean matrix factorization (which finds discrete {0,1} structure with automatic rank selection) with a logistic matrix factorization (which learns row and column biases capturing how link-prone each topic and material is). The two are fused as $\tilde{T}_{\mathrm{final}} = \sigma(\hat{T} + b_r + b_c)$, where $\hat{T}$ is the Boolean reconstruction, $b_r$ and $b_c$ are bias vectors, and $\sigma$ is the logistic sigmoid. This single equation converts discrete topic–material structure into calibrated link probabilities, and those probabilities are what the leave-out evaluation ranks.

What would settle it

Compile an independent list of superconducting transition-metal dichalcogenides from measured experimental data (for example, reported critical temperatures) that is not derived from the same topic clusters, mask those links, retrain the ensemble, and check whether hit@k and the score gap survive; if independently labeled superconductors are not ranked above non-superconductors, the claimed recovery is an artifact of the upstream clustering rather than a genuine material-property signal.

Watch

Extended reading notes

Core claim

The paper's central claim is that an ensemble of Boolean matrix factorization and logistic matrix factorization can rediscover withheld material–property links from a topic–material matrix built purely from text. On a corpus of 46,862 documents covering transition-metal dichalcogenides, hierarchical nonnegative matrix factorization produces a three-level topic tree; each material is linked to the topics found in its papers. The authors mask all entries linking four known superconductors (NbSe2, MoS2, S2Ta, Se2Ta) to the superconductivity topic, train the ensemble on the masked matrix, and report hit@3 = 1.0 for all four compounds, hit@1 = 1.0 for one, and a clean separation of scores: superconductors receive posterior scores around 0.70–0.81 while chemically similar non-superconductors receive at most 0.206. They conclude that the method distinguishes true superconductors from similar compounds and can prioritize candidates for experimental validation.

Load-bearing premise

The argument assumes that a material's link to a 'superconductivity' topic in the HNMFk-derived Materials Property Matrix is a faithful ground-truth marker that the material actually superconducts; if that link is not a reliable proxy for the physical property, the leave-out recovery shows only that the ensemble can reproduce its own clustering.

Editorial extensions

If this is right

If the recovery result is correct, the pipeline can be run on the unlabelled remainder of the TMD family and produce a ranked shortlist of candidate superconductors for experimental synthesis.
The same masked-link protocol can be applied to any property that leaves a textual trace, such as energy storage or tribology, without building a new supervised model for each property.
Because each predicted link sits inside a three-level topic tree, the prediction carries an explanation: the material is linked to a specific research theme, which an expert can inspect before acting.
The score separation implies that a single global threshold near 0.5 can screen a material family, separating positives from negatives without per-material tuning.
The approach is corpus-agnostic, so it transfers to other material families or any incomplete relational dataset with sparse binary links.

Reading between the lines

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

Because the masked 'positive' links are entries of a topic-derived matrix rather than independent experimental labels, the near-perfect hit rates may partly reflect the ensemble's ability to reconstruct its own upstream clustering; testing against an externally curated database of superconductivity would separate those two explanations.
The same BNMFk + LMF recipe applies naturally to other sparse binary relation graphs, such as drug–target or disease–gene networks, where Boolean interpretability and calibrated probabilities are both desirable; the paper does not report such experiments.
A stronger quantitative test would correlate the predicted link scores with a continuous physical quantity such as measured superconducting transition temperature $T_c$; the paper only reports binary separation, so the ranking's physical resolution remains untested.
Equation (3)'s additive fusion of a Boolean reconstruction with logistic biases is a general design pattern; replacing the sigmoid with another calibration function or weighting the two terms is a cheap ablation that the paper leaves open.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 paper is a legitimate application of your group's matrix factorization toolbox to a materials literature corpus, and the Table 1 leave-out on eleven compounds is a real prediction, not just matrix reconstruction. But the headline 'near-perfect top-3 precision' is about recovering HNMFk's own cluster assignments, so it doesn't validate material-property discovery. That's the thing to know.

What's actually new: combining HNMFk, BNMFk-LMF, and a Streamlit dashboard into a TMD link-prediction benchmark. The components are prior work (HNMFk [7], BNMFk-LMF [1], BUNIE [15]), and the paper is honest about that. The integration is clean and the dashboard sounds useful. The leave-out protocol is described clearly, and the fully masked ranking in Section 4.4 is genuinely predictive: the model never sees a positive label for those 11 compounds, yet the four known superconductors land on top. That is the strongest part of the paper.

The soft spots are real but not fatal. The hit@k test in Section 4.2 masks entries of the Materials Property Matrix, which HNMFk itself generated from the corpus. So hit@1 > 0.88 and hit@3 = 1.00 show the ensemble can reconstruct its upstream clustering, not that it finds superconductors. To make the discovery claim stick, the masked positives need to come from an independent source, e.g., experimentally labeled superconductors used as test labels. The negatives have the same problem: zeros are sampled as non-superconductors, but a zero only means no detected association in the text, not that the material lacks the property. That likely inflates the score separation in Figure 3. Also, there are no baselines at all. A random predictor, a co-occurrence count, or a standard link prediction method would put the numbers in context. Table 1 has no variance or significance. The counts are inconsistent: 73 TMDs in the abstract, 72 in Section 4.1; the conclusion says a 15-compound test set, but Section 4.4 lists 11. Small errors, but they undermine confidence in the pipeline's bookkeeping.

If I were refereeing this, I'd ask for external validation, baselines, and a cleaned data/code release. As a literature-mining tool, the approach is plausible and the dashboard is a practical contribution. As 'material property discovery,' the claim is stronger than the evidence. Still, it is a solid application paper that deserves referee time.

Referee Report

4 major / 6 minor

Summary. This manuscript describes a literature-based link-prediction pipeline for transition-metal dichalcogenide (TMD) materials. The authors use BUNIE to select a 46,862-document corpus, HNMFk to construct a three-level topic hierarchy and a binary Materials Property Matrix (topics x materials), and a BNMFk-LMF ensemble to score missing topic-material links. They validate the approach by masking 'verified superconducting links' for four known superconductors, then measuring hit@1/hit@3, the score separation between masked positives and sampled zeros, and a ranking of eleven materials. The reported results are near-perfect top-3 recovery and a clear score separation, together with an interactive dashboard for human-in-the-loop exploration.

Significance. The paper's strengths are its reproducible code release (T-ELF), automatic model selection inside HNMFk, and a concrete human-in-the-loop artifact. If the validation were externally grounded, the approach would be a useful hypothesis-generation tool for materials science. However, the central evaluation currently tests whether the ensemble reconstructs the outputs of its own upstream HNMFk clustering; it does not test whether the method predicts experimentally confirmed material properties. The paper also lacks baseline comparisons and an external ground-truth test set, so the significance claim about 'material property discovery' is not yet supported. The proposed pipeline remains a plausible and interpretable framework, but its empirical claims need substantial re-validation.

major comments (4)

[§4.1, §4.2] The validation protocol masks 'verified superconducting links' that are entries of the Materials Property Matrix, which §4.1 defines from HNMFk topic assignments over the corpus. A one entry in this matrix means the material was placed in a topic cluster by HNMFk; it is not an independent experimental measurement of superconductivity. The leave-out experiment in §4.2 therefore measures the ensemble's ability to reconstruct the upstream clustering, not its ability to identify superconducting materials. To support the central claim, the authors should re-run the evaluation with positive and negative labels taken from an independent source (e.g., experimentally verified superconductor databases or explicit literature annotations) rather than from the matrix being completed.
[§4.2 (negative sampling), §4.1] Negatives are sampled from 'zero' entries of the Materials Property Matrix, but §4.1 states that a zero means 'no such association' and that a separate value (NaN) denotes 'insufficient information to determine the relationship.' A zero is therefore an absence of detected topic co-occurrence, not evidence that the material is a non-superconductor. Sampling these entries as hard negatives and showing that they receive lower scores than masked positives (Fig. 3) may simply separate 'topic co-occurrence' from 'no detected topic co-occurrence.' The analysis should restrict negatives to materials that are independently known to be non-superconducting, or at minimum treat unobserved entries as unknown rather than as negatives.
[§4.2 (evaluation), §3.3] No baseline method is compared under the same leave-out protocol. The claimed near-perfect hit@3 and clean score separation could be matched or exceeded by simpler approaches such as LMF alone, HNMFk topic co-occurrence counts, or even a degree-based heuristic, and the paper provides no ablation to show that the BNMFk-LMF ensemble in Eq. (3) contributes beyond its components. A baseline table with the same masking protocol is needed before the framework can be described as 'reliably' rediscovering hidden links.
[§4.4, Table 1, §5] The ranking experiment behind Table 1 is not described with the same detail as §4.2: it is unclear how the eleven compounds were selected, whether the 'fully masked' procedure differs from the four-compound masking, how many random splits were used, and what variance the scores carry. The conclusion states that the method ranked all superconductors above all non-superconductors in a '15-compound test set,' but Table 1 lists eleven compounds. These protocol and reporting gaps make the threshold claim (≈0.5) difficult to assess. Please provide an explicit protocol, confidence intervals for Table 1, and reconcile the count discrepancy.

minor comments (6)

[§1, §4.1] The number of known TMDs is inconsistent: 73 in the introduction and abstract, versus 'only 72 known compounds' and 72 matrix columns in §4.1.
[Abstract, §1, §3.3] The abstract has a typo ('thier' for 'their'), §1 has 'derived form' instead of 'derived from', and §3.3 has a grammatical fragment ('we use to construct').
[Fig. 2, §4.2] The numeric 95% confidence intervals are described as narrow but not reported; please include exact intervals and define what a 'random split' means in the three-fold repetition.
[Fig. 3] The caption's note that 'these are identical for the four violins' is ambiguous; clarify which data are identical and consider plotting the negative distribution once.
[Front matter] The manuscript uses a placeholder ACM conference template with a dummy DOI, a 2018 copyright notice, and 'Conference acronym 'XX'; this should be corrected for the target venue.
[Data availability / §4.1] No data availability statement is provided. Since the corpus and the generated Materials Property Matrix are central to reproducibility, please state where the corpus, the matrix, and the dashboard are available in addition to the T-ELF code.

Circularity Check

2 steps flagged · score 6.0 of 10

Validation positives are entries of the Materials Property Matrix produced by the same HNMFk pipeline, so hit@3 and score separation demonstrate internal consistency rather than external material-property prediction.

fitted input called prediction [Section 4.1-4.2]
"Using the TMDs corpus and the topics extracted via HNMFk, we construct the binary Materials Property Matrix. ... Each entry M_ij in the matrix is defined as follows: M_ij = 1 indicates that the jth material is associated with the ith topic, M_ij = 0 denotes no such association... To evaluate whether our BNMFk + LMF ensemble can rediscover true but withheld superconducting links, we perform a leave-out experiment... remove all verified superconducting links from the compound–superconductor cluster matrix, and randomly sample an equal number of non-superconducting (zero) entries as negatives."

The positives being 'rediscovered' are entries of the Materials Property Matrix, which the paper itself constructs from HNMFk topics over the same corpus. Masking and recovering those entries tests whether BNMFk-LMF can reconstruct the low-rank structure that HNMFk imposed on the matrix. The experimental facts that NbSe2, MoS2, S2Ta, and Se2Ta superconduct are used only to select the benchmark and to name the recovered links; they are not the labels being masked. Hence hit@3 = 1.000 demonstrates internal consistency with the upstream factorization, not a validated material-property prediction.
self definitional [Section 4.2-4.3]
"randomly sample an equal number of non-superconducting (zero) entries as negatives. ... Positive edges cluster tightly near 1 ... Negatives are strongly skewed toward zero..."

The 'non-superconducting' negatives are defined as zero entries of the Materials Property Matrix, where zero is defined as 'no association' between material and topic, not as experimentally non-superconducting; the matrix separately uses NaN for insufficient information. The clean score separation in Figure 3 may therefore separate 'co-occurs with superconductivity-related text in the corpus' from 'does not', rather than superconductors from non-superconductors. The claimed separation of classes is thus built into the definition of the labels being scored.

full rationale

The central validation is circular in a specific, quotable way: the 'verified superconducting links' and 'non-superconducting zero entries' are entries of the Materials Property Matrix that Section 4.1 defines from the HNMFk topic decomposition of the same corpus. The near-perfect hit@k and violin separation show that the BNMFk-LMF ensemble can reproduce the latent structure of its own upstream clustering. The external knowledge that the four benchmark compounds are real superconductors appears only in the choice of materials and the interpretation of results, not as a test label in the leave-out protocol. Self-citations (e.g., [1], [7]) are method citations to code-released prior work and are not themselves load-bearing for the circularity; the problem is the validation protocol, not the citations. An external benchmark with independently measured superconducting labels would be needed to support the paper's material-property claims.

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

No physical entities are invented; the only new construct is the Materials Property Matrix, a data-structure artifact rather than a postulated entity. The load-bearing assumptions are about whether text-derived topic assignments can stand in for real material properties and whether matrix zeros are true negatives.

free parameters (4)

HNMFk depth L = 3
Depth of three levels chosen by bootstrap stability analysis (Section 4.1); treated as fixed for the reported results.
NMFk/BNMFk rank k
Rank selected by the NMFk stability heuristic; the heuristic itself is a modeling choice, not an externally fixed value.
LMF regularization lambda
Regularization strength in the LMF objective (Equation 2) is not specified in the paper, yet results depend on it.
Global decision threshold = approximately 0.5
In Section 4.4 the authors state that a threshold around 0.5 separates positives and negatives; this threshold appears chosen after inspecting the scores.

assumptions (5)

domain assumption TF-IDF features of titles and abstracts are a sufficient semantic representation for grouping papers into material-property topics.
Used in Section 4.1 to construct the input matrix for HNMFk; if false, the topic tree and all downstream links are arbitrary.
domain assumption Co-occurrence of a material name and topic tokens in a document indicates a functional association between the material and the topic/property.
This is what makes the Materials Property Matrix meaningful; Section 4.1 defines M_ij=1 when material j is associated with topic i. No external validation is provided.
domain assumption Zero entries in the Materials Property Matrix mean the material does not have the property.
In Section 4.2 zeros are sampled as negatives, but the paper defines zero as no detected association, not as confirmed absence.
standard math The latent low-rank model captures true structure in the topic-material matrix.
LMF and BNMFk assume low-rank structure, a standard matrix-completion assumption inherited from prior work.
domain assumption HNMFk's stability-based model selection finds the true number of topics.
Automatic rank selection is used throughout without external topic-quality validation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Topic Modeling and Link-Prediction for Material Property Discovery." pith.science (2026). https://pith.science/paper/SP2F4BW7

@misc{pith2026250706139,
  author       = {Pith},
  title        = {Pith review of: Topic Modeling and Link-Prediction for Material Property Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SP2F4BW7}},
  note         = {Machine review of arXiv:2507.06139}
}

read the original abstract

Link prediction infers missing or future relations between graph nodes, based on connection patterns. Scientific literature networks and knowledge graphs are typically large, sparse, and noisy, and often contain missing links between entities. We present an AI-driven hierarchical link prediction framework that integrates matrix factorization to infer hidden associations and steer discovery in complex material domains. Our method combines Hierarchical Nonnegative Matrix Factorization (HNMFk) and Boolean matrix factorization (BNMFk) with automatic model selection, as well as Logistic matrix factorization (LMF), we use to construct a three-level topic tree from a 46,862-document corpus focused on 73 transition-metal dichalcogenides (TMDs). These materials are studied in a variety of physics fields with many current and potential applications. An ensemble BNMFk + LMF approach fuses discrete interpretability with probabilistic scoring. The resulting HNMFk clusters map each material onto coherent topics like superconductivity, energy storage, and tribology. Also, missing or weakly connected links are highlight between topics and materials, suggesting novel hypotheses for cross-disciplinary exploration. We validate our method by removing publications about superconductivity in well-known superconductors, and show the model predicts associations with the superconducting TMD clusters. This shows the method finds hidden connections in a graph of material to latent topic associations built from scientific literature, especially useful when examining a diverse corpus of scientific documents covering the same class of phenomena or materials but originating from distinct communities and perspectives. The inferred links generating new hypotheses, produced by our method, are exposed through an interactive Streamlit dashboard, designed for human-in-the-loop scientific discovery.

Figures

Figures reproduced from arXiv: 2507.06139 by the authors.

**Figure 1.** Dashboard used to explore the HNMFk hierarchy. 4.2 Top 𝑘 retrieval of hidden links To evaluate whether our BNMF𝑘 + LMF ensemble can rediscover true but withheld superconducting links, we perform a leave-out experiment on four benchmark transition-metal dichalcogenides: NbSe2, MoS2, S2Ta, and Se2Ta. (1) Masking. For each compound, remove all verified superconducting links from the compound–superconductor cluster ma… view at source ↗

**Figure 3.** Split-violin plot of posterior probabilities for 24 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 16 canonical work pages

[7]

Maksim E Eren, Manish Bhattarai, Robert J Joyce, Edward Raff, Charles Nicholas, and Boian S Alexandrov. 2023. Semi-supervised classification of malware families under extreme class imbalance via hierarchical non-negative matrix factorization with automatic model selection. ACM Transactions on Privacy and Security 26, 4 (2023), 1–27

work page 2023
[1]

Ryan Barron, Maksim E Eren, Duc P Truong, Cynthia Matuszek, James Wen- delberger, Mary F Dorn, and Boian Alexandrov. 2025. Matrix Factorization for Inferring Associations and Missing Links. arXiv preprint arXiv:2503.04680 (2025)

work page Pith review arXiv 2025
[15]

Nicholas Solovyev, Ryan Barron, Manish Bhattarai, Maksim E Eren, Kim Ø Rasmussen, and Boian S Alexandrov. 2023. Interactive distillation of large single- topic corpora of scientific papers. In 2023 International Conference on Machine Learning and Applications (ICMLA). IEEE, 1000–1005

work page 2023
[2]

Manish Bhattarai, Ryan Barron, Maksim Eren, Minh Vu, Vesselin Grantcharov, Ismael Boureima, Valentin Stanev, Cynthia Matuszek, Vladimir Valtchinov, Kim Rasmussen, and Boian Alexandrov. 2024. HEAL: Hierarchical Embed- ding Alignment Loss for Improved Retrieval and Representation Learning. arXiv:2412.04661 [cs.IR] https://arxiv.org/abs/2412.04661

arXiv 2024
[3]

Emmanuel J Candes and Yaniv Plan. 2010. Matrix completion with noise. Proc. IEEE 98, 6 (2010), 925–936

work page 2010
[4]

Bolun Chen, Fenfen Li, Senbo Chen, Ronglin Hu, and Ling Chen. 2017. Link prediction based on non-negative matrix factorization. PLOS ONE 12, 8 (2017), e0182968. https://doi.org/10.1371/journal.pone.0182968

doi:10.1371/journal.pone.0182968 2017
[5]

Daniel M Dunlavy, Tamara G Kolda, and Evrim Acar. 2011. Temporal link prediction using matrix and tensor factorizations.ACM Transactions on Knowledge Discovery from Data (TKDD) 5, 2 (2011), 1–27

work page 2011
[6]

Maksim Eren, Nick Solovyev, Ryan Barron, Manish Bhattarai, Duc Truong, Ismael Boureima, Erik Skau, Kim Rasmussen, and Boian Alexandrov. 2023. Tensor Extraction of Latent Features (T-ELF) . Technical Report. Los Alamos National Laboratories. https://doi.org/10.5281/zenodo.10257897

work page doi:10.5281/zenodo.10257897 2023

Show all 18 references

[8]

Sunny Gupta, Jun-Jie Zhang, Jincheng Lei, Henry Yu, Mingjie Liu, Xiaolong Zou, and Boris I Yakobson. 2025. Two-Dimensional Transition Metal Dichalcogenides: A Theory and Simulation Perspective. Chemical Reviews (2025)

2025
[9]

Seyed Mehran Kazemi and David Poole. 2018. Simple embedding for link predic- tion in knowledge graphs. Advances in neural information processing systems 31 (2018)

2018
[10]

Siyu Liu, Tongqi Wen, Beilin Ye, Zhuoyuan Li, Han Liu, Yang Ren, and D Srolovitz
[11]

Yong Liu, Min Wu, Chunyan Miao, Peilin Zhao, and Xiao-Li Li. 2016. Neigh- borhood regularized logistic matrix factorization for drug-target interaction prediction. PLoS computational biology 12, 2 (2016), e1004760

2016
[12]

Aditya Krishna Menon and Charles Elkan. 2011. Link prediction via matrix fac- torization. In Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2011, Athens, Greece, September 5-9, 2011, Proceedings, Part II 22. Springer, 437–452

2011
[13]

Ratha Pech, Dong Hao, Liming Pan, Hong Cheng, and Tao Zhou. 2017. Link prediction via matrix completion. Europhysics Letters 117, 3 (2017), 38002

2017
[14]

Siamak Ravanbakhsh, Barnabás Póczos, and Russell Greiner. 2016. Boolean matrix factorization and noisy completion via message passing. In International Conference on Machine Learning . PMLR, 945–954

2016
[16]

Jamshid Sourati and James A Evans. 2023. Accelerating science with human- aware artificial intelligence. Nature human behaviour 7, 10 (2023), 1682–1696

2023
[17]

Joren Van Herck, María Victoria Gil, Kevin Maik Jablonka, Alex Abrudan, Andy S Anker, Mehrdad Asgari, Ben Blaiszik, Antonio Buffo, Leander Choudhury, Clemence Corminboeuf, et al. 2025. Assessment of fine-tuned large language models for real-world chemistry and material science...

2025
[2025]

Digital Discovery (2025)

Large Language Models for Material Property Predictions: elastic constant tensor prediction and materials design. Digital Discovery (2025)

2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [7]

Maksim E Eren, Manish Bhattarai, Robert J Joyce, Edward Raff, Charles Nicholas, and Boian S Alexandrov. 2023. Semi-supervised classification of malware families under extreme class imbalance via hierarchical non-negative matrix factorization with automatic model selection. ACM Transactions on Privacy and Security 26, 4 (2023), 1–27

work page 2023

[2] [1]

Ryan Barron, Maksim E Eren, Duc P Truong, Cynthia Matuszek, James Wen- delberger, Mary F Dorn, and Boian Alexandrov. 2025. Matrix Factorization for Inferring Associations and Missing Links. arXiv preprint arXiv:2503.04680 (2025)

work page Pith review arXiv 2025

[3] [15]

Nicholas Solovyev, Ryan Barron, Manish Bhattarai, Maksim E Eren, Kim Ø Rasmussen, and Boian S Alexandrov. 2023. Interactive distillation of large single- topic corpora of scientific papers. In 2023 International Conference on Machine Learning and Applications (ICMLA). IEEE, 1000–1005

work page 2023

[4] [2]

Manish Bhattarai, Ryan Barron, Maksim Eren, Minh Vu, Vesselin Grantcharov, Ismael Boureima, Valentin Stanev, Cynthia Matuszek, Vladimir Valtchinov, Kim Rasmussen, and Boian Alexandrov. 2024. HEAL: Hierarchical Embed- ding Alignment Loss for Improved Retrieval and Representation Learning. arXiv:2412.04661 [cs.IR] https://arxiv.org/abs/2412.04661

arXiv 2024

[5] [3]

Emmanuel J Candes and Yaniv Plan. 2010. Matrix completion with noise. Proc. IEEE 98, 6 (2010), 925–936

work page 2010

[6] [4]

Bolun Chen, Fenfen Li, Senbo Chen, Ronglin Hu, and Ling Chen. 2017. Link prediction based on non-negative matrix factorization. PLOS ONE 12, 8 (2017), e0182968. https://doi.org/10.1371/journal.pone.0182968

doi:10.1371/journal.pone.0182968 2017

[7] [5]

Daniel M Dunlavy, Tamara G Kolda, and Evrim Acar. 2011. Temporal link prediction using matrix and tensor factorizations.ACM Transactions on Knowledge Discovery from Data (TKDD) 5, 2 (2011), 1–27

work page 2011

[8] [6]

Maksim Eren, Nick Solovyev, Ryan Barron, Manish Bhattarai, Duc Truong, Ismael Boureima, Erik Skau, Kim Rasmussen, and Boian Alexandrov. 2023. Tensor Extraction of Latent Features (T-ELF) . Technical Report. Los Alamos National Laboratories. https://doi.org/10.5281/zenodo.10257897

work page doi:10.5281/zenodo.10257897 2023