FDE: maths behind, fitting optionsHello,
I receive this question from acustomer while mentioning FDE:
" ... I gather we’re talking about GAMs? I’ve only ever used R for these, so would be good to know what capabilities JMP has.
The kind of questions I have are:
- I see that JMP supports B and P splines, wavelets and Fourier. Does it have things like thin plate splines or tensor product bases?
- I’m also curious about distributions - does it handle the full exponential family?
- What flexibility is there on defining complexity (e.g. smooth term verses number of knots)? How are smooth terms fitted?
- What model diagnostic tools are available?
"
For the last question we have the Residual plots, any thing else ? For all other questions/comments documentation/help would be appreciated. Thanks in advance.
Emmanuel
Re: FDE: maths behind, fitting options- I see that JMP supports B and P splines, wavelets and Fourier. Does it have things like thin plate splines or tensor product bases?
- No, what you see in FDE is what is available. Pro 17 introduced wavelets, which are often much better than the other existing options in terms of speed and accuracy.
- I’m also curious about distributions - does it handle the full exponential family?
- You can change the distribution in GenReg, but for FDE models we currently recommend staying with the default Gaussian option, since the FPCA scores are both positive and negative.
- What flexibility is there on defining complexity (e.g. smooth term verses number of knots)? How are smooth terms fitted?
- You can use BIC, AICc, or GCV to select the number of knots. I pretty much only use wavelets now, though, because the whole knot location/selection step is avoided entirely and you go straight to the FPCA dimension selection step.
- What model diagnostic tools are available?
- The main tool I use is actual by predicted and residual by predicted plots.
Re: FDE: maths behind, fitting optionsOverall, the user really seems to be asking about GAMs, which we don't really support, and are not really the same thing as a functional model in in the sense of the FDE platform.
Re: FDE: maths behind, fitting optionsI'm teaching our four half-day FDE course October 23-26 in afternoon CET, morning US time. There is a fee associated, but feel free to contact me for a code to give to your customer to waive the fee. The course goes into great detail on the models, model comparison, analysis, and interpretation.
https://www.jmp.com/en_us/training/courses/jmp-pro-analyzing-curves-and-profile-using-the-functional-data-explorer.html
--Di
Re: FDE: maths behind, fitting optionsThank you,
we have a meeting planned with the customer I will see if he needs more info and if he's interested on courses ... Thanks
Re: FDE: maths behind, fitting optionsHello @chris_gotwalt1 ,
the guy is head of Data Scientist and wants to understand FDE better before promoting its usage to Engineers (potential additional JMP Pro revenues). He is Statistician using R & Python but he's keen to promote JMP Pro to Engineers as far he is able to understand exactly how this is working and thus appropriately support them when needed. He doesn't need a demo but rather scientific publication / papers on the topic. Is that available somewhere (I had a look in Confluence but did not find it).
Note: they acquired 1 JMP Pro for FDE following Discovery and the workshop in London.
Emmanuel
Re: FDE: maths behind, fitting optionsI'm not Chris, but I did a lot of research while Mark and I developed the FDE course. There isn't one good paper out there. The book by Ramsey and Silverman, Functional Data Analysis, is excellent. I learned most of what I know about wavelet models from Nason, Wavelet Methods in Statistics with R. Has this company licensed (for free) our JMP Education course materials? Maybe he just needs to look at our course PDF. It shows very well "exactly how this is working." I have a wavelet supplement that is not included in the licensed materials that I can send as well.
