24451_en-US

liliconcom

Constrain the Profiler for PLS models

To avoid extrapolation the PLS Profiler (and from my point of view in general when dealing with correlated variables) should be constrained within the correlation structure of the variables. For making it simple I tried to add the Hotelling T2 equation as constraint script to the Profiler, but this does not work, because the Profiler only accepts simple constraints. As workaround I am currently using the Hotelling T2 equation as additional response in the model to ensure not making extrapolations pulling a kind of convex hull around the feasible model space. This point is really important, since most people are not even aware that they are doing extrapolation when not considering the correlation structure of the variables.

-silvio.silvio

Statistics and AnalyticsRe: Constrain the Profiler for PLS models - Status changed to: Investigating

@mjoner It looks like you can save the T square formula now.  Hope that helps.

Re: Constrain the Profiler for PLS models

@chris_gotwalt1, I see a Save T^2 but not a Save T^2 Formula. Am I missing something?

Re: Constrain the Profiler for PLS models

A simple approach that is worth trying is to save a T^2 formula and give it a minimize or match target desireability function in the Profiler platform along with the other preidction formulas that you are working with.   

Re: Constrain the Profiler for PLS models - Status changed to: Open for VotingRe: Constrain the Profiler for PLS models<meta http-equiv="Content-Type" content="text/html; charset=UTF-8" />

This was reiterated in Fit Model &gt; PLS &gt; Issue that PLS profiler currently relies on predictor independence where silvio.silvio suggested that this a more general problem for the Profiler:

Nevertheless, I would like to have the general capability of constraining Profilers when ever dealing with correlated predictors (e.g. GLR, tree based methods, Neural Nets....). Also for these methods there is a high risk of predicting points outside the model space without being aware of this.

Re: Constrain the Profiler for PLS models

Michael,

I know that there is an add in for PLS and that Alex is working on this. But this problem is also relevant for other modeling methods that can handle correlated predictors (e.g. GLR, tree based methods, NN...), it would be great if JMP could provide general solution for this.

Best Regards

Silvio

Re: Constrain the Profiler for PLS models

Silvio, have you talked to liang.zw (or maybe varbanov.ar ) about this? They've done a fair bit of mathematical work to allow us to constrain the profiler in a PLS model. I've written a JMP add-in for this. There are a few issues with the script (especially if you're using imputation) but I should be able to resolve those sometime after JMP 13 comes out.

Re: Constrain the Profiler for PLS models

Based on some quick conversations with others internal to P&G, we believe this feature has been delivered, both as red triangle option, but even better in terms of having Extrapolation Control profilers.

Re: Constrain the Profiler for PLS models - Status changed to: Delivered