"multivariate curve resolution"... a NMF using ALS?

question
statistics

#1

Dear all,

I am using Julia and the JuliaStats ecosystem together with SciKit-learn (thanks PyCall) to perform component analysis and blind-source separation in various Raman and also XAFS spectra of glasses with varying chemical compositions.

Reading papers, I learned that people call this field “Chemometrics”, and several commercial packages are available to do so (e.g., http://www.eigenvector.com/software/pls_toolbox.htm)… In particular, several authors refer to a technic called “multivariate curve resolution” (MCR), which aims at extracting endmembers components and associated concentrations from spectra of mixed chemical components.

I usually perform component extraction using the NMF package in JuliaStats, with using the ALS method in particular… From the descriptions in the papers (in particular, see http://www.sciencedirect.com/science/article/pii/S0169743903002077), I understood that MCR uses an ALS algorithm with a non-negative constrain… So its seems to me that MCR-ALS and NMF-ALS are similar…

I therefore wanted to know if people in the JuliaStats community have some know knowledge about MCR. Is it just a NMF performed using an ALS algorithm, or is it something different? Do somebody have any additional knowledge about Chemometrics? Is it just a name given to the application of usual stat/ML technics (PCA, PLS, NMF, ICA…) to spectroscopic/chemical data or is there additional things making Chemometrics specific?

Thanks in advance!

Best,
Charles.


#2

So I will reply to myself on this topic, in case someone is interested.

It turns out that the “MCR” technic is a NMF technic performed most of the time using an ALS algorithm, but with specific conditions: closure (sum of the component fractions = 1), uni-modality (components vary in a logical way along a chemical profil, for instance), and of course non-negativity.

So it is quite easy to use Julia NMF algorithm to implement such chemometric technic.

A good paper to read about that is

de Juan, A., and R. Tauler (2006), Multivariate Curve Resolution (MCR) from 2000: Progress in Concepts and Applications, Critical Reviews in Analytical Chemistry, 36(3–4), 163–176, doi:10.1080/10408340600970005.

Best,
C.