Functional Data Analysis and Classification: How to calculate FPC of new data (unseen by the original fit) using initial model's Fourier basis?

[#1]

Hi!

I am running an experiment using several sensors out of a laboratory experiment where I've been able to collect about 700 samples , each with about 200 points of functional data, I've been using Functional Data Explorer for dimensional reduction through fitting of a Fourier basis and then its corresponding FPCs.

I am following the example " Fermentation Process" example found in documentation (page 299 of JMP 16.2 manual), I will later run a classifier to identify class 0 or 1, which is already labeled in the same data. All good so far, but how can I test this classifier against unseen data? how can I obtain the corresponding FPCs out of new data using the original fourier fit??

Many thanks!!

Rual

[#2 · Accepted solution]

Hi @RiztCL,

The easiest way to use a predefined functional model for new data without modifying it by adding the new data inside would be to create a validation column and use it in the Functional Data Explorer dialog window.
When creating the validation column, you can then specify your original training data points as training data set, and the new data as test set.
By using this validation column in the modeling, JMP should only use the training dataset (you original training runs) to fit the functional Fourier model, and you could then extract the FPCs for both the training and test sets.

I hope this solution will work for you,

[#3]

Hi Victor, Many thanks for the reply , it is indeed a good idea However, that would force me to always pass through the functional explorer,  is there anyway to obtain a set of Formulas to directly calculate the Fourier base and/or the FPCs by just adding new rows to the table??

Thanks!

Raul

[#4]

Hi Raul,

I have checked quickly in the platform, since you can't save formulas for FPC's or for the model's predictions, I don't think there is a direct option to calculate Fourier base and/or FPCs when adding rows.
There might be a more elegant and less manual way to calculate these parameters for new rows by scripting it (but I'm not an expert in scripting).

[#5]

Hi Victor,

Ok, no problem, I will search for a workaround, I *think* the fourier coefficients as well as the random components tables can be used to reconstruct the fourier basis formula, but it probably will involve the 'coding' of the Fourier formula to make the basis .. same thing about FPCs....

Let's see, thanks anyway

Cheers

Raul