Core-Conditioned Regularized Matrix Tri-Factorization for High-Dimensional Structured Systems

Ronald Katende

arxiv: 2409.06321 · v2 · submitted 2024-09-10 · 🧮 math.NA · cs.NA

Core-Conditioned Regularized Matrix Tri-Factorization for High-Dimensional Structured Systems

Ronald Katende This is my paper

Pith reviewed 2026-05-23 21:02 UTC · model grok-4.3

classification 🧮 math.NA cs.NA

keywords matrix tri-factorizationregularized low-rank approximationcore conditioningalternating minimizationKurdyka-Łojasiewicz propertyperturbation bounds

0 comments

The pith

A regularized tri-factorization A ≈ PDQ lets the central core D be explicitly conditioned while proving convergence of alternating minimization under Kurdyka-Łojasiewicz assumptions.

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

The paper develops a matrix approximation framework that decomposes A into side factors P and Q around a central core D whose numerical conditioning can be penalized or constrained directly. It shows that, in the unweighted full-data quadratic case, exact alternating minimization produces a descent sequence whose iterates remain bounded and converge to a critical point whenever the objective satisfies the Kurdyka-Łojasiewicz property. The same analysis supplies existence of minimizers once the regularizer is coercive, product-level perturbation bounds, and well-posed block updates. Validation on noisy and ill-conditioned low-rank problems indicates that the method recovers competitive reconstruction error while returning diagnostics such as the learned core condition number that spectral baselines omit. The authors position the approach as complementary to truncated SVD, useful precisely when both approximation fidelity and factor conditioning must be managed together.

Core claim

In the full-data quadratic setting the regularized PDQ objective admits minimizers under coercive regularization; its alternating-minimization iterates descend, stay bounded, and converge to a critical point under the Kurdyka-Łojasiewicz property, while the formulation also yields explicit product-level perturbation bounds and block-system well-posedness.

What carries the argument

The PDQ tri-factorization with explicit regularization or constraint on the conditioning of the central core matrix D.

If this is right

Product-level perturbation bounds hold for the reconstructed matrix.
Block updates remain well-posed for the quadratic objective.
The learned core condition number becomes an available diagnostic alongside reconstruction error.
The method is not claimed to outperform randomized SVD on pure spectral compression speed.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same core-conditioning device could be inserted into other factorizations that already admit alternating-minimization schemes.
Diagnostic reporting of core condition number may help decide when a low-rank model should be rejected on numerical grounds before deployment.
The current weighted missing-entry implementation is reported as non-competitive, suggesting the full-data analysis does not automatically transfer to incomplete-data regimes.

Load-bearing premise

The regularization must be coercive and the objective function must satisfy the Kurdyka-Łojasiewicz property.

What would settle it

An explicit full-data quadratic instance in which a coercive regularizer is used yet the alternating-minimization iterates diverge or fail to reach a critical point.

read the original abstract

This paper studies a regularized matrix tri-factorization \(A\approx PDQ\), where \(P\) and \(Q\) are side factors and \(D\) is a central core whose conditioning can be explicitly regularized or constrained. The formulation is a structured low-rank approximation framework, not a replacement for LU, QR, Cholesky, or the singular value decomposition. In the unregularized full-data Frobenius rank-\(r\) problem, truncated SVD remains the optimal benchmark. The contribution here concerns the regularized and core-conditioned setting, where reconstruction accuracy is treated together with factor scale, numerical conditioning, perturbation behavior, and weighted approximation. The analysis establishes the algebraic scope of the \(PDQ\) representation, proves existence of minimizers under coercive regularization, identifies the non-uniqueness induced by latent-space transformations, derives well-posed block updates for the quadratic full-data objective, and gives product-level perturbation bounds. For exact alternating minimization in the full-data quadratic case, it proves descent, boundedness of iterates, and convergence to a critical point under standard Kurdyka--\L{}ojasiewicz assumptions. A full multi-seed validation indicates competitive behavior in noisy and ill-conditioned low-rank approximation while reporting diagnostics not provided by purely spectral baselines, including the learned core condition number and block-system conditioning. The validation also clarifies the method's limits: randomized SVD remains faster for pure spectral compression, and the current weighted missing-entry variant is not uniformly competitive with matrix-completion baselines. The framework is therefore best viewed as a regularized and diagnostically transparent tri-factorization for settings where approximation quality and numerical conditioning must be controlled jointly.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Tri-factorization with explicit core conditioning and some analysis, but convergence claim rests on an unverified KL assumption for this loss.

read the letter

The paper puts forward a regularized tri-factorization A ≈ PDQ that lets the user directly control the conditioning of the middle core D. That is the concrete addition over plain low-rank factorizations. It works through the quadratic full-data case by deriving block updates, proving existence under coercive regularization, and supplying product-level perturbation bounds. The validation runs on noisy and ill-conditioned examples and reports the learned core condition number plus block conditioning, which standard SVD baselines do not give. Those diagnostics are the practical plus. Randomized SVD stays faster for pure compression, and the weighted missing-data version does not beat standard matrix-completion methods, so the method is positioned for joint accuracy-plus-conditioning tasks rather than as a general replacement. The convergence statement for exact alternating minimization claims descent, bounded iterates, and convergence to a critical point under standard Kurdyka-Łojasiewicz assumptions plus coercivity. The text invokes these but does not check whether the specific regularized objective satisfies the KL inequality at its critical points, so the result stays conditional. That is the main soft spot; the rest of the algebraic scope and bounds look standard and correctly scoped. The work is for people who already need structured low-rank approximations in ill-conditioned numerical settings and want the extra conditioning knobs and diagnostics. It is coherent enough on its own terms to go to a serious referee, though the KL step will need explicit verification or a weaker convergence argument.

Referee Report

2 major / 0 minor

Summary. The manuscript introduces a core-conditioned regularized matrix tri-factorization A ≈ PDQ, where the central core D can have its conditioning explicitly regularized or constrained. It claims to establish the algebraic scope of the representation, prove existence of minimizers under coercive regularization, identify non-uniqueness from latent-space transformations, derive well-posed block updates for the quadratic full-data objective, give product-level perturbation bounds, and prove descent, boundedness, and convergence to a critical point for exact alternating minimization under standard Kurdyka-Łojasiewicz assumptions. Numerical validation on noisy and ill-conditioned low-rank approximation tasks reports competitive performance together with diagnostics such as learned core condition number and block-system conditioning.

Significance. If the claimed existence, perturbation, and convergence results hold, the framework supplies a diagnostically transparent alternative to purely spectral methods when approximation quality must be balanced against numerical conditioning and factor scale. The explicit reporting of core condition number and block conditioning is a practical strength not provided by truncated SVD baselines. The analysis is positioned as complementary rather than competitive with standard factorizations or matrix completion methods.

major comments (2)

[Abstract] Abstract: the convergence statement for exact alternating minimization (descent, bounded iterates, convergence to critical point) is established only under the Kurdyka-Łojasiewicz property plus coercivity of the regularizer. No derivation or verification is supplied that the specific regularized tri-factorization objective satisfies the KL inequality (or admits a suitable desingularizing function) at its critical points; the result therefore remains conditional on an external property whose validity for this loss is not confirmed.
[Abstract] Abstract: existence of minimizers is asserted under coercive regularization, yet the manuscript supplies neither the explicit form of the regularizer nor a proof that the chosen regularizer is coercive on the product space; this step is load-bearing for well-posedness.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on the analysis claims in the abstract. We respond point-by-point to the major comments below.

read point-by-point responses

Referee: [Abstract] Abstract: the convergence statement for exact alternating minimization (descent, bounded iterates, convergence to critical point) is established only under the Kurdyka-Łojasiewicz property plus coercivity of the regularizer. No derivation or verification is supplied that the specific regularized tri-factorization objective satisfies the KL inequality (or admits a suitable desingularizing function) at its critical points; the result therefore remains conditional on an external property whose validity for this loss is not confirmed.

Authors: We agree that the convergence result is stated under the Kurdyka-Łojasiewicz property without supplying a specific verification or desingularizing function for the tri-factorization objective. The manuscript presents the result as conditional on this standard assumption from the non-convex optimization literature. To address the concern, we will revise the abstract to more explicitly highlight the conditional nature of the convergence statement and add a short discussion in the convergence section on the applicability of the KL property to coercive regularized objectives of this form. revision: partial
Referee: [Abstract] Abstract: existence of minimizers is asserted under coercive regularization, yet the manuscript supplies neither the explicit form of the regularizer nor a proof that the chosen regularizer is coercive on the product space; this step is load-bearing for well-posedness.

Authors: We agree that the explicit form of the regularizer and a proof of its coercivity on the product space are essential to substantiate the existence claim and are currently insufficiently detailed. In the revised manuscript we will supply the explicit expression for the regularizer and include a complete proof of coercivity to support well-posedness. revision: yes

Circularity Check

0 steps flagged

No significant circularity; convergence stated conditionally on external KL property

full rationale

The paper's central claims concern existence of minimizers under coercive regularization, well-posed block updates, perturbation bounds, and convergence of alternating minimization to a critical point. All such statements are explicitly conditioned on standard external assumptions (coercivity and the Kurdyka-Łojasiewicz property) rather than derived from the paper's own fitted quantities or definitions. No equations, parameters, or self-citations are shown that reduce the claimed results to inputs by construction. The framework is presented as a regularized tri-factorization whose diagnostic properties are analyzed separately from the conditional convergence guarantee.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no explicit free parameters, axioms, or invented entities can be extracted or audited.

pith-pipeline@v0.9.0 · 5825 in / 1251 out tokens · 28881 ms · 2026-05-23T21:02:31.431609+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages

[1]

and Atia, G

Rahmani, M. and Atia, G. K., ”High Dimensional Low Rank Plus Sparse Matrix Decomposition,” IEEE Transactions on Signal Processing, vol. 65, no. 8, pp. 2004-2019 , 2017, doi: 10.1109/TSP.2017.2649482

work page doi:10.1109/tsp.2017.2649482 2004
[2]

and Bouman, C., ”Covariance Estimation for High Dimension al Data Vectors Using the Sparse Matrix Transform,” in Advances in Neural Information Processing S ystems, vol

Cao, G. and Bouman, C., ”Covariance Estimation for High Dimension al Data Vectors Using the Sparse Matrix Transform,” in Advances in Neural Information Processing S ystems, vol. 21, 2008. [Online]. Avail- able: https://proceedings.neurips.cc/ paperfiles/paper/2008/ﬁle/8b16ebc056e613024c057be590b542eb- Paper.pdf

work page 2008
[3]

and Atia, G

Rahmani, M. and Atia, G. K., ”Randomized subspace learning appro ach for high dimensional low rank plus sparse matrix decomposition,” in Proc. 2015 49th Asilomar Conferen ce on Signals, Systems and Computers, pp. 1796-1800, 2015, doi: 10.1109/ACSSC.2015.7421461

work page doi:10.1109/acssc.2015.7421461 2015
[4]

R., and Bouman, C

Cao, G., Bachega, L. R., and Bouman, C. A., ”The Sparse Matrix Tr ansform for Covariance Estimation and Analysis of High Dimensional Signals,” IEEE Transactions on Image Processing, vol. 20, no. 3, pp. 625-640, 2011, doi: 10.1109/TIP.2010.2071390

work page doi:10.1109/tip.2010.2071390 2011
[5]

Li, Y., ”Sparse machine learning models in bioinformatics,” Electronic Theses and Dissertations, 2014, paper 5023

work page 2014
[6]

25, 2012

Ibrahimi, M., Javanmard, A., and Roy, B., ”Eﬃcient Reinforcement Learning for High Dimensional Linear Quadratic Systems,” in Advances in Neural Information Processing Systems, vol. 25, 2012. [Online]. Avail- able: https://proceedings.neurips.cc/ paperfiles/paper/2012/ﬁle/a9eb812238f753132652ae09963a05e9- Paper.pdf

work page 2012
[7]

Luo, X., Zhou, Y., Liu, Z., and Zhou, M., ”Fast and Accurate Non-N egative Latent Factor Analysis of High-Dimensional and Sparse Matrices in Recommender Systems,” IE EE Transactions on Knowledge and Data Engineering, vol. 35, no. 4, pp. 3897-3911, 2023, doi: 10.110 9/TKDE.2021.3125252

work page arXiv 2023
[8]

Jiang, J., Li, W., Dong, A., Gou, Q., and Luo, X., ”A Fast Deep AutoEn coder for High-Dimensional and Sparse Matrices in Recommender Systems,” Neurocomputing, v ol. 412, pp. 381-391, 2020, doi: https://doi.org/10.1016/j.neucom.2020.06.109

work page doi:10.1016/j.neucom.2020.06.109 2020
[9]

Du, K.-L., Swamy, M. N. S., Wang, Z.-Q., and Mow, W. H., ”Matrix Fact orization Techniques in Ma- chine Learning, Signal Processing, and Statistics,” Mathematics, v ol. 11, no. 12, art. 2674, 2023, doi: 10.3390/math11122674

work page doi:10.3390/math11122674 2023
[10]

Manzhos, S., and Ihara, M., ”Advanced Machine Learning Metho ds for Learning from Sparse Data in High- Dimensional Spaces: A Perspective on Uses in the Upstream of Deve lopment of Novel Energy Technologies,” Physchem, vol. 2, no. 2, pp. 72-95, 2022, doi: 10.3390/physchem 2020006

work page doi:10.3390/physchem 2022
[11]

Chen, J., Yang, S., Wang, Z., and Mao, H., ”Eﬃcient Sparse Repre sentation for Learning with High- Dimensional Data,” IEEE Transactions on Neural Networks and Lea rning Systems, vol. 34, no. 8, pp. 4208-4222, 2023, doi: 10.1109/TNNLS.2021.3119278

work page doi:10.1109/tnnls.2021.3119278 2023
[12]

K., ”A Nonlinear Matrix Decomposition for Mining the Zeros of Sparse Data,” SIAM Journal on Mathematics of Data Science, vol

Saul, L. K., ”A Nonlinear Matrix Decomposition for Mining the Zeros of Sparse Data,” SIAM Journal on Mathematics of Data Science, vol. 4, no. 2, pp. 431-463, 2022, do i: 10.1137/21M1405769. 6

work page doi:10.1137/21m1405769 2022
[13]

2020 IEEE Inte rnational Conference on Knowledge Graph (ICKG), pp

Wu, D., Lu, G., and Xu, Z., ”Robust and Accurate Representatio n Learning for High-Dimensional and Sparse Matrices in Recommender Systems,” in Proc. 2020 IEEE Inte rnational Conference on Knowledge Graph (ICKG), pp. 489-496, 2020, doi: 10.1109/ICBK50248.2020 .00075

work page doi:10.1109/icbk50248.2020 2020
[14]

125, art

Salehi, H., Gorodetsky, A., Solhmirzaei, R., and Jiao, P., ”High-Dime nsional Data Analytics in Civil Engineering: A Review on Matrix and Tensor Decomposition,” Engineer ing Applications of Artiﬁcial Intelligence, vol. 125, art. 106659, 2023, doi: https://doi.org/10 .1016/j.engappai.2023.106659

work page arXiv 2023
[15]

Roozbeh, M., Babaie-Kafaki, S., and Aminifard, Z., ”Improved Hig h-Dimensional Regression Models with Matrix Approximations Applied to the Comparative Case Studies with S upport Vector Machines,” Opti- mization Methods and Software, vol. 37, no. 5, pp. 1912-1929, 20 22, doi: 10.1080/10556788.2021.2022144

work page doi:10.1080/10556788.2021.2022144 1912
[16]

Y. Yuan, Q. He, X. Luo, and M. Shang, ”A Multilayered-and-Ran domized Latent Factor Model for High- Dimensional and Sparse Matrices,” IEEE Transactions on Big Data, v ol. 8, no. 3, pp. 784-794, 2022, doi: 10.1109/TBDATA.2020.2988778

work page doi:10.1109/tbdata.2020.2988778 2022
[17]

Z. Sun, G. Pedretti, P. Mannocci, E. Ambrosi, A. Bricalli, and D. I elmini, ”Time Complexity of In-Memory Solution of Linear Systems,” IEEE Transactions on Electron Devices , vol. 67, no. 7, pp. 2945-2951, 2020, doi: 10.1109/TED.2020.2992435

work page doi:10.1109/ted.2020.2992435 2020
[18]

X. Fu, N. Vervliet, L. De Lathauwer, K. Huang, and N. Gillis, ”Com puting Large-Scale Matrix and Tensor Decomposition With Structured Factors: A Uniﬁed Nonconvex Optim ization Perspective,” IEEE Signal Processing Magazine, vol. 37, no. 5, pp. 78-94, 2020, doi: 10.1109 /MSP.2020.3003544

work page arXiv 2020
[19]

Tian and Y

Y. Tian and Y. Zhang, ”A comprehensive survey on regularizatio n strategies in machine learning,” Infor- mation Fusion, vol. 80, pp. 146-166, 2022, doi: https://doi.org/10 .1016/j.inﬀus.2021.11.005

work page 2022
[20]

Wu and X

D. Wu and X. Luo, ”Robust Latent Factor Analysis for Precise R epresentation of High-Dimensional and Sparse Data,” IEEE/CAA Journal of Automatica Sinica, vol. 8, n o. 4, pp. 796-805, 2021, doi: 10.1109/JAS.2020.1003533

work page doi:10.1109/jas.2020.1003533 2021
[21]

Chouzenoux and V

E. Chouzenoux and V. Elvira, ”Sparse Graphical Linear Dynamic al Systems,” Journal of Machine Learning Research, vol. 25, no. 223, pp. 1-53, 2024, URL: http://jmlr.org /papers/v25/23-0878.html

work page 2024
[22]

D. Cai, J. Nagy, and Y. Xi, ”Fast Deterministic Approximation of S ymmetric Indeﬁnite Kernel Matrices with High Dimensional Datasets,” SIAM Journal on Matrix Analysis and Applications, vol. 43, no. 2, pp. 1003-1028, 2022, doi: 10.1137/21M1424627

work page doi:10.1137/21m1424627 2022
[23]

Maalouf, I

A. Maalouf, I. Jubran, and D. Feldman, ”Fast and Accurate Le ast-Mean-Squares Solvers for High Di- mensional Data,” IEEE Transactions on Pattern Analysis and Machin e Intelligence, vol. 44, no. 12, pp. 9977-9994, 2022, doi: 10.1109/TPAMI.2021.3139612

work page doi:10.1109/tpami.2021.3139612 2022
[24]

Wesel and K

F. Wesel and K. Batselier, ”Large-Scale Learning with Fourier F eatures and Tensor Decompo- sitions,” Advances in Neural Information Processing Systems, vol. 34, pp. 17543-17554, 2021, URL: https://proceedings.neurips.cc/ paperfiles/paper/2021/ﬁle/92a08bf918f44ccd961477be30023da1- Paper.pdf

work page 2021
[25]

Kontolati, D

K. Kontolati, D. Loukrezis, D. G. Giovanis, L. Vandanapu, and M . D. Shields, ”A survey of unsupervised learning methods for high-dimensional uncertainty quantiﬁcation in black-box-type problems,” Journal of Computational Physics, vol. 464, pp. 111313, 2022, doi: https:// doi.org/10.1016/j.jcp.2022.111313

work page doi:10.1016/j.jcp.2022.111313 2022
[26]

W. Min, X. Wan, T.-H. Chang, and S. Zhang, ”A Novel Sparse Gra ph-Regularized Singular Value Decom- position Model and Its Application to Genomic Data Analysis,” IEEE Tra nsactions on Neural Networks and Learning Systems, vol. 33, no. 8, pp. 3842-3856, 2022, doi: 1 0.1109/TNNLS.2021.3054635

work page arXiv 2022
[27]

G. Cai, J. Li, X. Liu, Z. Chen, and H. Zhang, ”Learning and Compr essing: Low-Rank Matrix Factor- ization for Deep Neural Network Compression,” Applied Sciences, vo l. 13, no. 4, pp. 2704, 2023, doi: https://doi.org/10.3390/app13042704

work page doi:10.3390/app13042704 2023
[28]

Efroni, S

Y. Efroni, S. Kakade, A. Krishnamurthy, and C. Zhang, ”Spar sity in Partially Controllable Linear Systems,” in Proceedings of the 39th International Conference on Machine L earning, pp. 5851-5860, 2022, URL: https://proceedings.mlr.press/v162/efroni22b.html

work page 2022
[29]

D. Wu, X. Luo, M. Shang, Y. He, G. Wang, and M. Zhou, ”A Deep L atent Factor Model for High- Dimensional and Sparse Matrices in Recommender Systems,” IEEE Tr ansactions on Systems, Man, and Cybernetics: Systems, vol. 51, no. 7, pp. 4285-4296, 2021, doi: 10.1109/TSMC.2019.2931393. 7

work page doi:10.1109/tsmc.2019.2931393 2021
[30]

J.-S. Yeom, J. J. Thiagarajan, A. Bhatele, G. Bronevetsky, a nd T. Kolev, ”Data-Driven Performance Modeling of Linear Solvers for Sparse Matrices,” in 2016 7th Interna tional Workshop on Performance Modeling, Benchmarking and Simulation of High Performance Compute r Systems (PMBS), pp. 32-42, 2016, doi: 10.1109/PMBS.2016.009

work page doi:10.1109/pmbs.2016.009 2016
[31]

Rahmani and G

M. Rahmani and G. Atia, ”A Subspace Learning Approach for Hig h Dimensional Matrix Decomposition with Eﬃcient Column/Row Sampling,” in Proceedings of The 33rd Intern ational Conference on Machine Learning, pp. 1206-1214, 2016, URL: https://proceedings.mlr.pr ess/v48/rahmani16.html. 8

work page 2016

[1] [1]

and Atia, G

Rahmani, M. and Atia, G. K., ”High Dimensional Low Rank Plus Sparse Matrix Decomposition,” IEEE Transactions on Signal Processing, vol. 65, no. 8, pp. 2004-2019 , 2017, doi: 10.1109/TSP.2017.2649482

work page doi:10.1109/tsp.2017.2649482 2004

[2] [2]

and Bouman, C., ”Covariance Estimation for High Dimension al Data Vectors Using the Sparse Matrix Transform,” in Advances in Neural Information Processing S ystems, vol

Cao, G. and Bouman, C., ”Covariance Estimation for High Dimension al Data Vectors Using the Sparse Matrix Transform,” in Advances in Neural Information Processing S ystems, vol. 21, 2008. [Online]. Avail- able: https://proceedings.neurips.cc/ paperfiles/paper/2008/ﬁle/8b16ebc056e613024c057be590b542eb- Paper.pdf

work page 2008

[3] [3]

and Atia, G

Rahmani, M. and Atia, G. K., ”Randomized subspace learning appro ach for high dimensional low rank plus sparse matrix decomposition,” in Proc. 2015 49th Asilomar Conferen ce on Signals, Systems and Computers, pp. 1796-1800, 2015, doi: 10.1109/ACSSC.2015.7421461

work page doi:10.1109/acssc.2015.7421461 2015

[4] [4]

R., and Bouman, C

Cao, G., Bachega, L. R., and Bouman, C. A., ”The Sparse Matrix Tr ansform for Covariance Estimation and Analysis of High Dimensional Signals,” IEEE Transactions on Image Processing, vol. 20, no. 3, pp. 625-640, 2011, doi: 10.1109/TIP.2010.2071390

work page doi:10.1109/tip.2010.2071390 2011

[5] [5]

Li, Y., ”Sparse machine learning models in bioinformatics,” Electronic Theses and Dissertations, 2014, paper 5023

work page 2014

[6] [6]

25, 2012

Ibrahimi, M., Javanmard, A., and Roy, B., ”Eﬃcient Reinforcement Learning for High Dimensional Linear Quadratic Systems,” in Advances in Neural Information Processing Systems, vol. 25, 2012. [Online]. Avail- able: https://proceedings.neurips.cc/ paperfiles/paper/2012/ﬁle/a9eb812238f753132652ae09963a05e9- Paper.pdf

work page 2012

[7] [7]

Luo, X., Zhou, Y., Liu, Z., and Zhou, M., ”Fast and Accurate Non-N egative Latent Factor Analysis of High-Dimensional and Sparse Matrices in Recommender Systems,” IE EE Transactions on Knowledge and Data Engineering, vol. 35, no. 4, pp. 3897-3911, 2023, doi: 10.110 9/TKDE.2021.3125252

work page arXiv 2023

[8] [8]

Jiang, J., Li, W., Dong, A., Gou, Q., and Luo, X., ”A Fast Deep AutoEn coder for High-Dimensional and Sparse Matrices in Recommender Systems,” Neurocomputing, v ol. 412, pp. 381-391, 2020, doi: https://doi.org/10.1016/j.neucom.2020.06.109

work page doi:10.1016/j.neucom.2020.06.109 2020

[9] [9]

Du, K.-L., Swamy, M. N. S., Wang, Z.-Q., and Mow, W. H., ”Matrix Fact orization Techniques in Ma- chine Learning, Signal Processing, and Statistics,” Mathematics, v ol. 11, no. 12, art. 2674, 2023, doi: 10.3390/math11122674

work page doi:10.3390/math11122674 2023

[10] [10]

Manzhos, S., and Ihara, M., ”Advanced Machine Learning Metho ds for Learning from Sparse Data in High- Dimensional Spaces: A Perspective on Uses in the Upstream of Deve lopment of Novel Energy Technologies,” Physchem, vol. 2, no. 2, pp. 72-95, 2022, doi: 10.3390/physchem 2020006

work page doi:10.3390/physchem 2022

[11] [11]

Chen, J., Yang, S., Wang, Z., and Mao, H., ”Eﬃcient Sparse Repre sentation for Learning with High- Dimensional Data,” IEEE Transactions on Neural Networks and Lea rning Systems, vol. 34, no. 8, pp. 4208-4222, 2023, doi: 10.1109/TNNLS.2021.3119278

work page doi:10.1109/tnnls.2021.3119278 2023

[12] [12]

K., ”A Nonlinear Matrix Decomposition for Mining the Zeros of Sparse Data,” SIAM Journal on Mathematics of Data Science, vol

Saul, L. K., ”A Nonlinear Matrix Decomposition for Mining the Zeros of Sparse Data,” SIAM Journal on Mathematics of Data Science, vol. 4, no. 2, pp. 431-463, 2022, do i: 10.1137/21M1405769. 6

work page doi:10.1137/21m1405769 2022

[13] [13]

2020 IEEE Inte rnational Conference on Knowledge Graph (ICKG), pp

Wu, D., Lu, G., and Xu, Z., ”Robust and Accurate Representatio n Learning for High-Dimensional and Sparse Matrices in Recommender Systems,” in Proc. 2020 IEEE Inte rnational Conference on Knowledge Graph (ICKG), pp. 489-496, 2020, doi: 10.1109/ICBK50248.2020 .00075

work page doi:10.1109/icbk50248.2020 2020

[14] [14]

125, art

Salehi, H., Gorodetsky, A., Solhmirzaei, R., and Jiao, P., ”High-Dime nsional Data Analytics in Civil Engineering: A Review on Matrix and Tensor Decomposition,” Engineer ing Applications of Artiﬁcial Intelligence, vol. 125, art. 106659, 2023, doi: https://doi.org/10 .1016/j.engappai.2023.106659

work page arXiv 2023

[15] [15]

Roozbeh, M., Babaie-Kafaki, S., and Aminifard, Z., ”Improved Hig h-Dimensional Regression Models with Matrix Approximations Applied to the Comparative Case Studies with S upport Vector Machines,” Opti- mization Methods and Software, vol. 37, no. 5, pp. 1912-1929, 20 22, doi: 10.1080/10556788.2021.2022144

work page doi:10.1080/10556788.2021.2022144 1912

[16] [16]

Y. Yuan, Q. He, X. Luo, and M. Shang, ”A Multilayered-and-Ran domized Latent Factor Model for High- Dimensional and Sparse Matrices,” IEEE Transactions on Big Data, v ol. 8, no. 3, pp. 784-794, 2022, doi: 10.1109/TBDATA.2020.2988778

work page doi:10.1109/tbdata.2020.2988778 2022

[17] [17]

Z. Sun, G. Pedretti, P. Mannocci, E. Ambrosi, A. Bricalli, and D. I elmini, ”Time Complexity of In-Memory Solution of Linear Systems,” IEEE Transactions on Electron Devices , vol. 67, no. 7, pp. 2945-2951, 2020, doi: 10.1109/TED.2020.2992435

work page doi:10.1109/ted.2020.2992435 2020

[18] [18]

X. Fu, N. Vervliet, L. De Lathauwer, K. Huang, and N. Gillis, ”Com puting Large-Scale Matrix and Tensor Decomposition With Structured Factors: A Uniﬁed Nonconvex Optim ization Perspective,” IEEE Signal Processing Magazine, vol. 37, no. 5, pp. 78-94, 2020, doi: 10.1109 /MSP.2020.3003544

work page arXiv 2020

[19] [19]

Tian and Y

Y. Tian and Y. Zhang, ”A comprehensive survey on regularizatio n strategies in machine learning,” Infor- mation Fusion, vol. 80, pp. 146-166, 2022, doi: https://doi.org/10 .1016/j.inﬀus.2021.11.005

work page 2022

[20] [20]

Wu and X

D. Wu and X. Luo, ”Robust Latent Factor Analysis for Precise R epresentation of High-Dimensional and Sparse Data,” IEEE/CAA Journal of Automatica Sinica, vol. 8, n o. 4, pp. 796-805, 2021, doi: 10.1109/JAS.2020.1003533

work page doi:10.1109/jas.2020.1003533 2021

[21] [21]

Chouzenoux and V

E. Chouzenoux and V. Elvira, ”Sparse Graphical Linear Dynamic al Systems,” Journal of Machine Learning Research, vol. 25, no. 223, pp. 1-53, 2024, URL: http://jmlr.org /papers/v25/23-0878.html

work page 2024

[22] [22]

D. Cai, J. Nagy, and Y. Xi, ”Fast Deterministic Approximation of S ymmetric Indeﬁnite Kernel Matrices with High Dimensional Datasets,” SIAM Journal on Matrix Analysis and Applications, vol. 43, no. 2, pp. 1003-1028, 2022, doi: 10.1137/21M1424627

work page doi:10.1137/21m1424627 2022

[23] [23]

Maalouf, I

A. Maalouf, I. Jubran, and D. Feldman, ”Fast and Accurate Le ast-Mean-Squares Solvers for High Di- mensional Data,” IEEE Transactions on Pattern Analysis and Machin e Intelligence, vol. 44, no. 12, pp. 9977-9994, 2022, doi: 10.1109/TPAMI.2021.3139612

work page doi:10.1109/tpami.2021.3139612 2022

[24] [24]

Wesel and K

F. Wesel and K. Batselier, ”Large-Scale Learning with Fourier F eatures and Tensor Decompo- sitions,” Advances in Neural Information Processing Systems, vol. 34, pp. 17543-17554, 2021, URL: https://proceedings.neurips.cc/ paperfiles/paper/2021/ﬁle/92a08bf918f44ccd961477be30023da1- Paper.pdf

work page 2021

[25] [25]

Kontolati, D

K. Kontolati, D. Loukrezis, D. G. Giovanis, L. Vandanapu, and M . D. Shields, ”A survey of unsupervised learning methods for high-dimensional uncertainty quantiﬁcation in black-box-type problems,” Journal of Computational Physics, vol. 464, pp. 111313, 2022, doi: https:// doi.org/10.1016/j.jcp.2022.111313

work page doi:10.1016/j.jcp.2022.111313 2022

[26] [26]

W. Min, X. Wan, T.-H. Chang, and S. Zhang, ”A Novel Sparse Gra ph-Regularized Singular Value Decom- position Model and Its Application to Genomic Data Analysis,” IEEE Tra nsactions on Neural Networks and Learning Systems, vol. 33, no. 8, pp. 3842-3856, 2022, doi: 1 0.1109/TNNLS.2021.3054635

work page arXiv 2022

[27] [27]

G. Cai, J. Li, X. Liu, Z. Chen, and H. Zhang, ”Learning and Compr essing: Low-Rank Matrix Factor- ization for Deep Neural Network Compression,” Applied Sciences, vo l. 13, no. 4, pp. 2704, 2023, doi: https://doi.org/10.3390/app13042704

work page doi:10.3390/app13042704 2023

[28] [28]

Efroni, S

Y. Efroni, S. Kakade, A. Krishnamurthy, and C. Zhang, ”Spar sity in Partially Controllable Linear Systems,” in Proceedings of the 39th International Conference on Machine L earning, pp. 5851-5860, 2022, URL: https://proceedings.mlr.press/v162/efroni22b.html

work page 2022

[29] [29]

D. Wu, X. Luo, M. Shang, Y. He, G. Wang, and M. Zhou, ”A Deep L atent Factor Model for High- Dimensional and Sparse Matrices in Recommender Systems,” IEEE Tr ansactions on Systems, Man, and Cybernetics: Systems, vol. 51, no. 7, pp. 4285-4296, 2021, doi: 10.1109/TSMC.2019.2931393. 7

work page doi:10.1109/tsmc.2019.2931393 2021

[30] [30]

J.-S. Yeom, J. J. Thiagarajan, A. Bhatele, G. Bronevetsky, a nd T. Kolev, ”Data-Driven Performance Modeling of Linear Solvers for Sparse Matrices,” in 2016 7th Interna tional Workshop on Performance Modeling, Benchmarking and Simulation of High Performance Compute r Systems (PMBS), pp. 32-42, 2016, doi: 10.1109/PMBS.2016.009

work page doi:10.1109/pmbs.2016.009 2016

[31] [31]

Rahmani and G

M. Rahmani and G. Atia, ”A Subspace Learning Approach for Hig h Dimensional Matrix Decomposition with Eﬃcient Column/Row Sampling,” in Proceedings of The 33rd Intern ational Conference on Machine Learning, pp. 1206-1214, 2016, URL: https://proceedings.mlr.pr ess/v48/rahmani16.html. 8

work page 2016