Using FDE and DOE to Help Build Predictive Models for Spectral Data (2020-EU-30MP-410)Level: Intermediate
Bill Worley, JMP Senior Global Enablement Engineer, SAS
In the recent past Partial Least Squares (PLS) has been used to build predictive models for spectral data. A newer approach using Functional Data Explorer (FDE) and Covariate Design of Experiments (DOE) will be shown that will allow for fewer spectra to be used in the development of a good predictive model. This method uses one-fourth to one-third of the data that would otherwise be used to build a predictive model based on spectral data.
August 3, 2020: An Advanced Mastering JMP video on this topic is now available.
BreakoutDOEPredictive ModelingRe: Using FDE and DOE to Help Build Predictive Models for Spectral Data (2020-EU-30MP-410)Hi. I just added a link to the video at the top of the page.
Re: Using FDE and DOE to Help Build Predictive Models for Spectral Data (2020-EU-30MP-410)Hi Lu,
I thought I was going to record this for JMP on Air at one point, but ended up not doing it. I would be happy to walk through it with you if that is of interest.
Best,
Bill
Re: Using FDE and DOE to Help Build Predictive Models for Spectral Data (2020-EU-30MP-410)Hey Bill,
Were you able to record this session and if yes, are you willing to share it with me?
Regards,
Lu
Re: Using FDE and DOE to Help Build Predictive Models for Spectral Data (2020-EU-30MP-410)Hi Bill, thanks for clarification this afternoon. I have tried to reproduce your FDE results using P-splines however I get other Y FPC scores, see attached file in columns Y FPC redo. Why is this difference?
Re: Using FDE and DOE to Help Build Predictive Models for Spectral Data (2020-EU-30MP-410)Hello, in the Functional Data Explorere Gasoline data set I tried several models however I can't reproduce the reported 4 Y FPC's; what model is used? All the 60 samples rows are used? Thanks for info, Frank
Re: Using FDE and DOE to Help Build Predictive Models for Spectral Data (2020-EU-30MP-410)Hi @frankderuyck and @martindemel,
I just saw this link and I am happy to record my session. I will work with Customer Care to get that out.
I will answer some of the individual questions from past posts.
For the FDE work you will need to work with Rows as Functions, and not the Stacked format. Place all wavelength columns in the Y, Output and use Sample Number (should be categorical) in the ID, Function. For those that have JMP Pro 15 you can put Octane in the Z, Supplementary box and ultimately use the Functional DOE to build a Generalized Regression model.
Hope this helps. Please come back with more questions as needed.
Best,
Bill
Re: Using FDE and DOE to Help Build Predictive Models for Spectral Data (2020-EU-30MP-410)Is a very useful topic, I would also appreciate to hear the recorded version
Re: Using FDE and DOE to Help Build Predictive Models for Spectral Data (2020-EU-30MP-410)@frankderuyck Under Graph menu there is a parallel plot Bill used. So you do not need to use the Graph Builder's "Parallel Plot" graph element.
Re: Using FDE and DOE to Help Build Predictive Models for Spectral Data (2020-EU-30MP-410)@Bill_Worley Would you be able to record your presentation and add it to this post? I believe this is of huge interest to lots of people 
Re: Using FDE and DOE to Help Build Predictive Models for Spectral Data (2020-EU-30MP-410)Thanks Martin, problem is that I cant activate parallel plot in Graphbuider? I have jmp 14.1 pro
Re: Using FDE and DOE to Help Build Predictive Models for Spectral Data (2020-EU-30MP-410)Hey Frank,
as far I can see from the attached gasoline data, he used the Parallel Plot, scaled uniformely and as Y al the NIR columns, to create the plot, not the graph builder. There you do not have an option to scale unitformely. This you may can first create scaled columns for all the NIR columnsdo beforehand , then the parallel plot in Graph Builder might provide similar result.
Martin
Re: Using FDE and DOE to Help Build Predictive Models for Spectral Data (2020-EU-30MP-410)Hello Bill,
very nice & useful presentation, pity you could not present it in Munich, I have some questions with regard to the strategy & analysis. I believe that you started with a PCA on you NIR-data and reduced the spectral data set to 4 principle components? Next steps are not clear to me: how did you organize the data so that they could be processed by the functional data explorer? Did you stack the wavelengths and the principle components? I guess finally you set up a main model custom DOE with the four retrieved FPCA's entered as covariate factors? As for the parallel plot, was this done with the graph builder?
These are my first questions, thanks in advance for reply!
Kind regards, Frank
