Talk:Partial least squares regression

Moving this article to Partial Least Squares
I think this article should be moved to "Partial Least Squares" for three reasons.
 * 1. That's what I've always heard it stands for.
 * 2. The SAS link refers to it that way.
 * 3. Google returns the usual large number of results for that title, and only 3 pages for the current title.

I'll change it in the next few days unless there is comment here on reasons not to. Spalding 11:16, May 9, 2005 (UTC)


 * I would strongly disagree with that. The style on Wikipedia is to minimize the use of capitals. The only articles I saw having all words starting with capitals are for people, and for book titles.


 * If you change this, you would also need to change least squares, and linear least squares and linear regression. I think this would be unacceptable. Wonder what others think. Oleg Alexandrov 16:12, 9 May 2005 (UTC)
 * What I meant was Partial least squares, replacing linear with least. I wasn't thinking about capitalization when I typed that.  I think the only other article that would have to be changed if we rename this article is the one or two that link here.  Spalding 19:32, May 9, 2005 (UTC)
 * Good point. Yes, I like the "least" thing more, and is more used. Oleg Alexandrov 19:53, 9 May 2005 (UTC)

Technical
Please add a paragraph explaining what it is/how it works without using "math jargon" (e.g. "hyperplanes of maximum variance"). Something for a first year University student (beginning calculus) would probably be about the level to aim for. -- 22:45, 20 December 2006 (UTC)
 * 18 mths on that hasn't happened. I too came here looking for information and just found jargon. Message to the folks who write these pages : If the average reader knew all the fancy words, he/she wouldn't be coming to Wikipedia for the answers. So as the guy above says, aim the article at an inquisitive reasonably-intelligent beginner level. Otherwise the exercise is pointless. Rcbutcher (talk) 09:11, 16 August 2008 (UTC)


 * Not to belabor this point, but this page is totally incomprehensible for the lay person. Wikipedia should read like an encyclopedia, not a graduate level math text.

Obviously it should have said "minimum variance"; I've fixed it. In one sense this is a relatively low-priority topic. The article suffers various other deficiencies as well, which would have been corrected otherwise. Michael Hardy (talk) 17:30, 16 August 2008 (UTC)

...and also: If it can be understood by a mathematician who's never heard of the method but not by a beginning calculus student, I wouldn't call it pointless. Michael Hardy (talk) 17:32, 16 August 2008 (UTC)

Is the pseudo-code in line (7) correct? It is "normalizing" by the squared norm of $t$, but the text says that $T$ is an orthonormal matrix, so its columns must have unit norm, and thus we should normalize by the non-squared Euclidean norm of $t$. --Jfessler (talk) 19:12, 23 January 2023 (UTC)

Name of the method?
The method seems to be more commonly called partial least squares in some communities than the alternative projection to latent structure so I have made a modification accordingly. Neither my edit nor the one it altered has a good citation. In my case, I am modifying it based on the standard in play in neuroimaging where it is routinely called partial least squares it is not clear what one would reference for something like this. 128.200.240.175 23:09, 26 July 2007 (UTC)

This page or reference is lacking a large quantity of information
We just ended a conference on PLS 9 Sept 2007. 

This is a major and large field of experimental, theoretical and applied statistics/mathematics. Then upon request, I was asked to check this page and I realize there is a great deal more to be said in about this. Just some momentary thoughts:


 * 1) Aka. "PLS" and a number of variations: LPLS, PLS-PM, PLS..etc etc...
 * 2) Missing initial and serious references on the topic
 * 3) Mathematical background with some simple derivations and examples
 * 4) more more...

—Preceding unsigned comment added by Jahibadkaret (talk • contribs) 17:40, 11 September 2007 (UTC)

It's hardly surprising that a lot is missing. Go ahead and add to the article. Michael Hardy 18:08, 11 September 2007 (UTC)

It is now 2010 ... still much room for improvement. PLS is more widely applied than ever; I'm surprised someone has not added to this article. I came here looking for info, and as was disappointed. I could add some, but I'm not a statistician. Anyone know how to put in diacritics? Need to add for Wold reference that I inserted. Janopus (talk) 11:19, 22 May 2009 (UTC)

I quite agree. I have made some small changes to bring it more up to date, added or updated a few references (including removal of a broken link and another to a professorial home page), but it could do with a major expansion. PLS path modelling needs an article on its own. Galltywenallt (talk) 14:54, 16 April 2011 (UTC)

described algorithm wrong ?
The statement "T is orthonormal matrix " is wrong, as T is a vector ! I implemented the algorithm and it does not work, maybe I made a mistake. Can anybody give a reference to its origin ? There are lots of PLS algorithms, but I've nevern seen this one. Found another error: If I start the PLS1 algorithm with y = (-1,0,1) then, T1 = (1/3, 1/3, 1/3) and q1 = 0 which breaks the loop indepenednt of X. Further: T0 is initalized, but T1 is never constructed, it just appears in line 5. —Preceding unsigned comment added by Rocksportrocker (talk • contribs) 15:02, 28 November 2008 (UTC) Wrong or not, I find the matrix algebra used by Wold easier to understand than this "pseudocode". I personally would replace it, but prefer to defer to a expert, particularly an expert practitioner. Inserting the code to me is over specialized; better to give a reference and instead just mention that various algorithms exist. Janopus (talk) 12:11, 22 May 2009 (UTC)

has anyone validated the content ?
Adding to the previous comment, i'm quite surprised the article mention that X=Y.B+B0, shouldn't it be Y=X.B+B0 ? —Preceding unsigned comment added by 82.225.154.2 (talk) 23:23, 21 May 2009 (UTC)
 * It definitely needs more work. Shyamal (talk) 04:28, 22 May 2009 (UTC)

The described algorithm does not specify how to compute W(1). I guess it's X(0)^T * y ...? Is there any reference for this one version of the algorithm .... ? —Preceding unsigned comment added by 82.225.154.2 (talk) 01:08, 2 June 2009 (UTC)

The loop starts with k=1 (which is weird, but why not), so i guess on lines 2&3 it should be X(1) and T(1) —Preceding unsigned comment added by 82.225.154.2 (talk) 01:12, 2 June 2009 (UTC)

Akila
Akila —Preceding unsigned comment added by 122.164.59.184 (talk) 06:18, 15 December 2009 (UTC)

Deleted commentary from this page
Yesterday someone deleted a section from this page. I replace it here in full: I notice that the IP address that deleted this section Geolocates to the University of Texas at Dallas. I also note that Prof Abdi works at the University of Texas at Dallas. Perhaps it was accidentally deleted but it seems in violation of the Wikipedia policy on talk pages (see WP:TPG). --Amaher (talk) 23:17, 26 January 2010 (UTC)


 * Congrats on replacing this. It is old and derives from a Revision as of 17:04, 18 October 2006 and had no userr-id or date associated. I think the point has been dealt with or no longer applies. But looking at the one remaining "Abdi" reference does raise the question of why 6 citations are required for a single point. My guess is that this "Abdi" reference discusses a lot more different things, that may already be in this article or not, that would be worth mentioning and citing appropriately. Melcombe (talk)

Merge subsection "Partial Least Squares SVD" with "PLS Correlation"?
The subsection "Partial Least Squares SVD" is fairly vague and points to a single reference: "Susceptibility of brain atrophy to TRIB3 [...]". This reference is only tangentially related to PLS: it applies the method without new methodological development. For this reason, it should be complemented with statistical papers that actually describe PLS-SVD. More importantly, upon reading the reference, I saw that the "PLS-SVD" used there is nothing else than PLS correlation (PLSC). This is clearly stated towards the end in the section "Statistical discovery": "Among the several PLS versions proposed in the literature, we focus on the symmetric formulation of PLS computed through the singular value decomposition of the cross-covariance matrix."

Accordingly, I propose to: 1) Move the content of the subsection "Partial Least Squares SVD" to "PLS Correlation". 2) Delete the subsection "Partial Least Squares SVD". 3) Add relevant references to the methodology of PLSC. For example, "Partial Least Squares (PLS) methods for neuroimaging: A tutorial and review" and other items from the references list. Nakatasatoru (talk) 20:59, 29 November 2023 (UTC)

Are the equations in the "underlying model" section complete?
It doesn't state how the two equations listed are related to each other. Later it does state that X and Y are decomposed such that there is maximum covariance between U and T, but it doesn't seem like that statement is enough. I feel like it should also include an equation like U = RT, with a definition for R. 140.141.192.46 (talk) 01:44, 8 February 2024 (UTC)