Pith. sign in

REVIEW 3 major objections 5 minor 22 references

Sparse Infrared Spectroscopy for Detection of Volatile Organic Compounds

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · 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 non-negative matrix factorization can design a sparse filter array of 20-50 infrared pass bands that identifies volatile organic compounds at 1-10 parts per million, replacing the 1,000+ spectral points of a…

desk verdict Useful design study, but the headline channel-count and detection-limit numbers are in-sample training accuracies and need an out-of-sample test or reframing before they are sold as hardware-relevant. read the letter →

arxiv 2506.20678 v1 pith:KZJTOIOC submitted 2025-06-22 physics.chem-ph physics.data-an

Mira Welner , Andre Hazbun , Thomas Beechem This is my paper · ORCID
classification physics.chem-phphysics.data-an
keywords sparseinfraredspectroscopynon-negativematrixfactorizationvolatileorganiccompoundsfilter-basedgasdetectionabsorptiontask-specificsensingvirtualexperiment
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 sets out to show that infrared spectroscopy for a narrow task does not need to collect a full spectrum. Using non-negative matrix factorization on a quantitative infrared absorption database of 40 volatile organic compounds, the authors identify a small set of spectral pass bands that act as measurement channels. Their virtual experiment reports that 20-50 such channels suffice to identify any of the 40 compounds at concentrations of 1-10 parts per million, compared with more than 1,000 spectral samples in a conventional measurement. If true, the result would let a fixed filter array replace a dispersive or interferometric spectrometer for a predefined set of gases, reducing size, weight, power, and cost without sacrificing detection performance.

What carries the argument

The carrying mechanism is non-negative matrix factorization (NMF): the non-negative absorption database $V$ is approximated as $V \approx S H$, where $H$ holds a small number of non-negative basis spectra and $S$ holds the per-channel scores. The paper's move is to treat each basis spectrum as the transmittance curve of a physical filter, so a detector reading is $S = H A$ for an unknown absorption $A$, and reconstruction $A_R = H^+ S$ uses the pseudo-inverse of the filter matrix. Sparsity arises because most of the infrared absorption data sits in a few broad bands, so NMF concentrates information into a few channels. The argument then reduces the spectroscopy task to choosing $\delta$ (20-50) and verifying by simulation that the nearest-database-match to $A_R$ identifies the compound at the target concentrations and noise levels.

What would settle it

Build a filter array whose transmittance curves are the NMF basis spectra from this paper, expose it to the 40 VOCs at 1-10 ppm in nitrogen, and check whether nearest-match identification with 20-50 channels reproduces the claimed accuracy; a concrete failure sign would be needing substantially more channels once filter manufacturing tolerances and correlated, non-Gaussian noise are included.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the basis spectra obtained from non-negative matrix factorization of a VOC absorption database can be interpreted directly as optical filter transmittance functions, making the compressed representation physical. A detector behind each such filter integrates the absorption spectrum against that basis, yielding one score per channel; multiplying the score vector by the pseudo-inverse of the basis matrix reconstructs the spectrum well enough to identify the compound by nearest-match to the database. The virtual experiment reports 100 percent identification accuracy with fewer than 50 channels down to roughly 2 $\mu$mol/mol carrier-gas concentration at a simulated noise level of about 0.2 percent, with detection extending to about 1 $\mu$mol/mol for the compound most similar to the database mean. In mixtures, a secondary contaminant can be detected down to roughly 10 percent of a known carrier concentration with 40-50 channels at 97 percent accuracy, a regime where uniformly spaced spectral sampling fails even with hundreds of channels. The paper also reports that performance extrapolates to concentrations not included in the matrix factorization design.

Load-bearing premise

The argument depends on NMF basis spectra being physically realizable as filter transmittance functions and on real measurement noise resembling the proportional, Gaussian error used in simulation; if either fails, a real 20-50 channel device may not reach the claimed detection limits.

Editorial extensions

If this is right

  • A detector array with tens of channels, not thousands, can identify a predefined set of VOCs at low parts-per-million concentrations, which would let portable infrared sensors replace lab-scale spectrometers for fixed monitoring tasks.
  • The same filter design works when the carrier gas is changed among compounds with different spectral variability, with similar channel counts and detection limits, suggesting the approach is not tuned to one background.
  • Because the approach interpolates to concentrations not used in the design, a single filter set could cover a continuous concentration range rather than requiring calibration at every level.
  • Standard uniform spectral sampling needs roughly three times as many channels and has a roughly four times worse detection limit in the virtual experiment, so the gap between task-specific and general-purpose sampling is large.
  • For a known VOC stream, contamination down to about 10 percent of the carrier concentration can be flagged with 40-50 channels at 97 percent accuracy, which is a concrete industrial leak or process-monitoring use case.

Reading between the lines

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

  • This design recipe is not limited to gas-phase infrared sensing: any modality with a non-negative database of target responses and a tunable filter array, such as Raman or UV-visible spectroscopy or electronic nose arrays, could use the same basis-as-filter construction, though channel counts and noise tolerance would need retesting.
  • The closed-set nature of the database means SIRS identifies only the 40 compounds it was trained on; detecting an unknown or out-of-list compound would require an additional rejection rule, which the paper does not provide.
  • A hardware-aware variant that optimizes filters subject to manufacturing constraints, such as achievable transmittance shapes, coating tolerances, and stray light, would be the natural next test; the current virtual experiment assumes the NMF basis is realizable exactly.
  • Because NMF bases are non-orthogonal and data-driven, they resemble compressive-sensing measurement matrices, and combining SIRS with compressed-reconstruction theory might provide rigorous guarantees on how few channels suffice for a given compound set.
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.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes Sparse Infrared Spectroscopy (SIRS), a filter-array sensing scheme in which non-negative matrix factorization (NMF) of a synthetic infrared absorption database is used to design a small number of spectral pass bands. The central claim is that with roughly 20-50 filter channels the approach can identify any of 40 volatile organic compounds at concentrations of 1-10 parts per million, compared with the more than 1,000 spectral samples used in conventional FTIR. The virtual experiment uses the NIST quantitative infrared database to build a training set at 16 concentrations, applies NMF to obtain basis spectra, reconstructs noisy spectra via the pseudo-inverse of the filter matrix, and identifies compounds by nearest-neighbor matching to the database. The paper also reports performance in a known carrier-gas mixture for three case-study compounds and under varying noise levels.

Significance. If the claimed channel reduction transfers to hardware, the approach would be a useful contribution to task-specific infrared sensing, since it decouples spectroscopic performance from instrument complexity. The strengths of the manuscript are the clear and internally consistent mathematical pipeline (NMF, pseudo-inverse reconstruction, nearest-neighbor classification), the use of a quantitative public database, and the explicit test of interpolation along the concentration axis by holding out selected concentrations. However, the headline numbers are obtained in an in-sample evaluation: the same 40 compounds and the same spectral library are used both to design the filters and to score identification. The significance of the 20-50 channel and 1-10 ppm claims therefore depends on demonstrating out-of-sample transfer, either to held-out compounds or to independent measurements.

major comments (3)
  1. [§III, Fig. 4] The central numerical claims are in-sample estimates. The NMF filter set H is learned from the synthetic database V built from the NIST absorption coefficients of the same 40 VOCs at 16 concentrations (§II), and identification is performed by reconstructing noisy spectra with Eq. (4) and comparing them to that same database (§III). The open markers in Fig. 4 are concentrations not used in the filter design, so they validate interpolation along the concentration axis only; they do not test whether the filters generalize to a different compound, a different instrument, or a different measurement condition. The abstract's '1-10 ppm' and '20-50 spectral samples' statements are therefore training-set accuracies, and the >1000-to-20-50 reduction may be an artifact of fitting and testing on the same library. Please add an out-of-sample split at the compound level, or at least an explicit statement that the reported numbers apply only to the 40-compound library used for design.
  2. [§II, Eqs. (2)-(5); §III, Fig. 5] The hardware implication in the abstract is not yet supported because the virtual experiment assumes that each optical filter has transmittance exactly equal to a row of H (Eq. 2) and that measurement noise is an independent multiplicative Gaussian perturbation (Eq. 5). No tolerance analysis is provided for deviations of the realized filter transmittance from the NMF basis, nor for stray light or correlated noise. The sensitivity of the detection limit to the noise parameter is in fact substantial: Fig. 5(a) reports a detection limit of 8 µmol/mol at 1% noise versus <2 µmol/mol at 0.2% noise. Please report how the channel counts and detection limits respond to perturbations of H and to more realistic noise, or temper the hardware claims accordingly.
  3. [§III, Fig. 4] The procedure used to obtain the 'minimum number of channels' is not specified, and the reported values have no uncertainty estimates. The manuscript does not state the search rule for the minimum channel count, the number of noise realizations used at each channel count, or the repeatability of the NMF solution across random initializations (the solver is explicitly random, §II). Since the central quantitative results are channel counts and detection limits, please provide this procedural detail and report variability over repeated NMF fits and noise draws.
minor comments (5)
  1. [§II] The text states that the synthetic database has 40×16 = 625 spectra; the correct product is 640. Please correct the arithmetic.
  2. [Fig. 4 caption and §III] The text above Fig. 4 describes the noise level as 0.02%, while the Fig. 4 caption states 0.2%. Please reconcile this discrepancy.
  3. [Fig. 2 caption] The Fig. 2(d) caption refers to 'bromoethane', whereas the text and Fig. 2(a) identify the compound as bromomethane (CH3Br). Please correct the label.
  4. [Fig. 4 and Eqs. (2)-(4)] The comparison against 'standard spectral sampling' using uniformly spaced Gaussian pass bands does not state the filter width or placement criteria, which makes the reported factor-of-three comparison difficult to reproduce. Please specify the baseline design.
  5. [§VI] The data availability statement says that data are available from the corresponding author upon reasonable request, but no code or processed data are deposited. Given the data-driven nature of the claims, making the NIST subset and analysis scripts available would materially aid reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline 20–50 channel and 1–10 ppm figures are in-sample estimates: the NMF-derived filter pass bands and the identification reference library come from the same synthetic NIST database.

  1. fitted input called prediction [Section II (Method and Approach) and Section III (Results and Discussion), Eqs. 1–5, Fig. 4]
    "The resulting database (V) is made up of R rows corresponding to each of the 3526 spectral energies sampled in the original dataset and C columns tabulating the absorption at each of these energies for the 40×16 = 625 separate simulated spectra. ... The noisy absorption was then used to simulate an effective score spectra using Eq. 3, where H was determined via the NMF process outlined above. Eq. 4 then reconstructed the spectrum, which was compared to the synthetic database."

    The filter pass bands are the NMF basis H, fit to the synthetic database V. The identification test then generates test spectra from the same 40 VOCs and the same Beer's-law model (Eq. 1), projects them through the fitted H, reconstructs with H+, and matches the result to the nearest column of the same V used for the fit. Thus the reported minimum channel counts and detection limits are in-sample training-set accuracies. The open-marker 'blind' concentrations in Fig. 4 are interpolations along the concentration axis of the same compounds and same spectral model, not independent measurements or unseen compounds.

full rationale

The paper contains no load-bearing self-citation and no imported uniqueness theorem; the NMF methodology is standard and the virtual experiment is transparently described. However, the central quantitative claim — that roughly 20–50 spectral samples identify VOCs at 1–10 ppm — is supported only by an in-sample evaluation: the NMF basis that defines the filters and the synthetic database against which reconstructed spectra are matched are the same V built from the same 40 NIST compounds and concentrations. The only held-out aspect is concentration interpolation, which remains within the same compound set and the same simulated noise model. Therefore the headline result is a training-set consistency result rather than an out-of-sample prediction. A proper external benchmark (unseen compounds, measured FTIR spectra, or hardware validation) is needed before the 20–50 channel reduction can be accepted as a property of the sensing scheme. This is a partial circularity in the evaluation protocol, not a full definitional collapse, so the score is 6.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities. The free parameters are simulation choices that affect the quantitative results. The axioms are standard physics and domain assumptions about the database, filter realizability, the classification rule, and the noise model; the filter realizability and noise model are the most speculative and directly affect the hardware implication.

free parameters (3)
  • Path length L = 20 m
    Chosen as typical of commercial long-path gas cells; directly scales absorption and therefore the reported parts-per-million detection limits.
  • Noise level P = 0 to 0.01 (0 to 1 percent)
    Simulation parameter in Eq. 5; the 0.2 percent and 1 percent cases drive the reported channel counts and detection limits.
  • Concentration sampling grid = 16 values from 0.1 to 100 umol/mol
    The grid defines the training and blind test concentrations; the interpolation claim depends on this density.
assumptions (5)
  • domain assumption Beer's Law as used in Eq. 1
    The relation log10(A) = -L sum alpha c is written with A as absorption, but the form implies A is transmittance; either way the simulation is internally consistent, but the convention is unstated.
  • domain assumption NIST Quantitative Infrared Database is quantitatively accurate for the 40 VOCs over 576 to 3974 cm^-1
    The entire virtual experiment rests on this database; no independent validation of the absorption coefficients is provided.
  • domain assumption NMF basis spectra can be physically realized as optical filters
    The hardware claim depends on filters whose transmittance equals the learned basis, which is not demonstrated.
  • domain assumption Minimum mean square difference to the synthetic database identifies the correct compound
    The classification rule in the virtual experiment is a standard nearest-neighbor approach, but it presumes the database contains the true source and that reconstruction error is a valid similarity metric.
  • domain assumption Multiplicative Gaussian noise models real FTIR measurement noise
    Real noise includes correlated baseline drift, detector nonlinearity, and temperature effects; the simple model in Eq. 5 is unvalidated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sparse Infrared Spectroscopy for Detection of Volatile Organic Compounds." pith.science (2026). https://pith.science/paper/KZJTOIOC

@misc{pith2026250620678,
  author       = {Pith},
  title        = {Pith review of: Sparse Infrared Spectroscopy for Detection of Volatile Organic Compounds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KZJTOIOC}},
  note         = {Machine review of arXiv:2506.20678}
}
abstract

To reduce the complexity of infrared spectroscopy hardware while maintaining performance, a data informed, task-specific, spectral collection approach termed Sparse Infrared Spectroscopy (SIRS) is developed. Using a numerically based virtual experiment based on a quantitatively accurate infrared database, non-negative matrix factorization is used to identify the spectral pass bands of a minimal number of filters necessary to identify volatile organic compounds (VOC) within either an inert background or mixture of gases. The data-driven approach is found capable of identifying contaminants at the 1-10 part per million level (PPM) with $\mathrm{\sim~20-50}$ spectral samples as opposed to the more than 1,000 typical of a traditional infrared spectrum. Reasonably robust to both noise and the characteristics of the base compound in a mixture, the task-specific spectral sampling points to simplified hardware design that maintains performance.

Figures

Figures reproduced from arXiv: 2506.20678 by the authors.

Figure 1
Figure 1. FIG. 1. Conceptual representation of task-specific sparse infrared spectroscopy (SIRS). A broad [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. (a) Simulated absorption for the 40 separate volatile organic compounds (VOC) when [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. (a) Evolution of reconstructed spectrum of acetone with channel number. (b) The evolution [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: FIG. 4. (a) Minimum number of channels necessary to correctly identify each VOC within a [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. (a) Required channels needed to identify a secondary component in a known acetone [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 17 canonical work pages

  1. [1]

    Freitag , author M

    author author S. Freitag , author M. Sulyok , author N. Logan , author C. T. \ Elliott ,\ and\ author R. Krska ,\ title title The potential and applicability of infrared spectroscopic methods for the rapid screening and routine analysis of mycotoxins in food crops , \ https://doi.org/10.1111/1541-4337.13054 journal journal Compr. Rev. Food Sci. Food Saf. ...

    work page arXiv

  2. [2]

    author author S. N. \ Zulkifli , author H. A. \ Rahim ,\ and\ author W.-J. \ Lau ,\ title title Detection of contaminants in water supply: A review on state-of-the-art monitoring technologies and their applications , \ https://doi.org/10.1016/j.snb.2017.09.078 journal journal Sensors and Actuators B: Chemical \ volume 255 ,\ pages 2657--2689 ( year 2018 ) NoStop

    work page doi:10.1016/j.snb.2017.09.078

  3. [3]

    author author M. S. \ Malachowski , author S. P. \ Levine , author G. Herrin , author R. C. \ Spear , author M. Yost ,\ and\ author Z. Yi ,\ title title Workplace and Environmental Air Contaminant Concentrations Measured by Open Path Fourier Transform Infrared Spectroscopy : A Statistical Process Control Technique to Detect Changes from Normal Operating C...

    work page arXiv

  4. [4]

    Rosenthal , author W

    author author P. Rosenthal , author W. Aarts , author A. Bonanno , author D. Boning , author S. Charpenay , author A. Gower , author M. Richter , author T. Smith , author P. Solomon , author M. Spartz , author C. Nelson , author A. Waldhauer , author J. Xu , author V. Yakovlev , author W. Zhang , author L. Allen , author B. Cordts , author M. Brandt , aut...

    work page

  5. [5]

    author author K. B. \ Be \'c , author J. Grabska ,\ and\ author C. W. \ Huck ,\ title title Principles and Applications of Miniaturized Near-Infrared ( NIR ) Spectrometers , \ https://doi.org/10.1002/chem.202002838 journal journal Chem. -- Eur. J. \ volume 27 ,\ pages 1514--1532 ( year 2021 ) NoStop

    doi:10.1002/chem.202002838

  6. [6]

    \ Sung , author R.-H

    author author L.-Y. \ Sung , author R.-H. \ Shie ,\ and\ author C.-J. \ Lu ,\ title title Locating sources of hazardous gas emissions using dual pollution rose plots and open path Fourier transform infrared spectroscopy , \ https://doi.org/10.1016/j.jhazmat.2013.11.006 journal journal Journal of Hazardous Materials \ volume 265 ,\ pages 30--40 ( year 2014...

    work page doi:10.1016/j.jhazmat.2013.11.006

  7. [7]

    \ Sung \ and\ author C.-J

    author author L.-Y. \ Sung \ and\ author C.-J. \ Lu ,\ title title A single-beam titration method for the quantification of open-path Fourier transform infrared spectroscopy , \ https://doi.org/10.1016/j.jqsrt.2014.04.016 journal journal Journal of Quantitative Spectroscopy and Radiative Transfer \ volume 145 ,\ pages 43--49 ( year 2014 ) NoStop

    work page doi:10.1016/j.jqsrt.2014.04.016 2014

  8. [8]

    \ Tsao , author P.-E

    author author Y.-C. \ Tsao , author P.-E. \ Chang , author S.-Y. \ Chen ,\ and\ author Y.-H. \ Hwang ,\ title title Real- Time Fab-Wise Airborne Molecular Contaminant ( AMC ) Monitoring System Using Multiple Fourier Transform Infrared ( FTIR ) Spectrometers in a Semiconductor Plant , \ https://doi.org/10.4209/aaqr.2014.08.0184 journal journal Aerosol Air ...

    work page doi:10.4209/aaqr.2014.08.0184

Show all 22 references
  1. [9]

    Cherniienko , author R

    author author A. Cherniienko , author R. Lesyk , author L. Zaprutko ,\ and\ author A. Pawe czyk ,\ title title IR-EcoSpectra : Exploring sustainable ex situ and in situ FTIR applications for green chemical and pharmaceutical analysis , \ https://doi.org/10.1016/j.jpha.2024.02....

    doi

  2. [10]

    Mortada , author Y

    author author B. Mortada , author Y. M. \ Sabry , author D. Khalil ,\ and\ author T. Bourouina ,\ title title Miniature Infrared Spectral Sensing Solutions for Ubiquitous Analytical Chemistry , \ in\ https://doi.org/10.1007/978-3-030-96534-1_13 booktitle Green Chemical Analysi...

    doi

  3. [11]

    author author J. H. \ Correia , author M. Bartek ,\ and\ author R. F. \ Wolffenbuttel ,\ title title Bulk-micromachined tunable Fabry -- Perot microinterferometer for the visible spectral range , \ https://doi.org/10.1016/S0924-4247(99)00023-0 journal journal Sensors and Actua...

    doi

  4. [12]

    author author M. C. M. M. \ Souza , author A. Grieco , author N. C. \ Frateschi ,\ and\ author Y. Fainman ,\ title title Fourier transform spectrometer on silicon with thermo-optic non-linearity and dispersion correction , \ https://doi.org/10.1038/s41467-018-03004-6 journal j...

    doi

  5. [13]

    Yang , author T

    author author Z. Yang , author T. Albrow-Owen , author W. Cai ,\ and\ author T. Hasan ,\ title title Miniaturization of optical spectrometers , \ https://doi.org/10.1126/science.abe0722 journal journal Science \ volume 371 ,\ pages eabe0722 ( year 2021 ) NoStop

    doi

  6. [14]

    Tittl , author A

    author author A. Tittl , author A. Leitis , author M. Liu , author F. Yesilkoy , author D.-Y. \ Choi , author D. N. \ Neshev , author Y. S. \ Kivshar ,\ and\ author H. Altug ,\ title title Imaging-based molecular barcoding with pixelated dielectric metasurfaces , \ https://doi...

    doi

  7. [15]

    Cerjan \ and\ author N

    author author B. Cerjan \ and\ author N. J. \ Halas ,\ title title Toward a Nanophotonic Nose : A Compressive Sensing-Enhanced , Optoelectronic Mid-Infrared Spectrometer , \ https://doi.org/10.1021/acsphotonics.8b01503 journal journal ACS Photonics \ volume 6 ,\ pages 79--86 (...

    doi

  8. [16]

    Oliver , author W

    author author J. Oliver , author W. Lee , author S. Park ,\ and\ author H.-N. \ Lee ,\ title title Improving resolution of miniature spectrometers by exploiting sparse nature of signals , \ https://doi.org/10.1364/OE.20.002613 journal journal Opt. Express, OE \ volume 20 ,\ pa...

    doi

  9. [17]

    Ni , author Q

    author author J. Ni , author Q. Zhang , author X. Zhang , author Z. Sun ,\ and\ author D. Bao ,\ title title Detection Method and Common Characteristics of Waste Solvent from Semiconductor Industry , \ https://doi.org/10.3390/molecules28165992 journal journal Molecules \ volum...

    doi

  10. [18]

    author author P. H. R. \ Ng , author S. Walker , author M. Tahtouh ,\ and\ author B. Reedy ,\ title title Detection of illicit substances in fingerprints by infrared spectral imaging , \ https://doi.org/10.1007/s00216-009-2806-9 journal journal Anal Bioanal Chem \ volume 394 ,...

    doi

  11. [19]

    author author P. M. \ Chu , author F. R. \ Guenther , author G. C. \ Rhoderick ,\ and\ author W. J. \ Lafferty ,\ title title The NIST quantitative infrared database , \ https://doi.org/10.6028/jres.104.004 journal journal J. Res. Natl. Inst. Stand. Technol. \ volume 104 ,\ pa...

    doi

  12. [20]

    Hecht ,\ @noop title Optics \ ( publisher Pearson Education ,\ year 2016 ) NoStop

    author author E. Hecht ,\ @noop title Optics \ ( publisher Pearson Education ,\ year 2016 ) NoStop

    2016

  13. [21]

    author author N. Gillis ,\ https://doi.org/10.48550/arXiv.1401.5226 title The Why and How of Nonnegative Matrix Factorization , \ ( year 2014 ),\ https://arxiv.org/abs/1401.5226 arXiv:1401.5226 [stat] NoStop

    2014 doi arXiv

  14. [22]

    Pedregosa , author F

    author author F. Pedregosa , author F. Pedregosa , author G. Varoquaux , author G. Varoquaux , author N. Org , author A. Gramfort , author A. Gramfort , author V. Michel , author V. Michel , author L. Fr , author B. Thirion , author B. Thirion , author O. Grisel , author O. Gr...

    2011

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.