Question on how to use Sequential Feature Selector

[Sebastian Khan]

Hi,

I am interested in using the "Sequential Feature Selector" (SFS) to fit some N-dimensional data with one output variable where N could be up to 7.

Is it correct that the I can use the mlxtend's SFS algorithms to find the "best" set of basis functions for each dimension (found by fitting against some training data and then compared against a validation set to prevent overfitting?)

I guess the first step is to learn how to use mlxtend to fit an N dimensional data set with a set a basis functions.

In my field one successful strategy to model the data I am working with is to model each dimension a separate 1D polynomial

and perform some tensor product between them to get the N-dimensional fit.

Is this something that can be done in mlxtend?

(I also posted this in gitter so sorry for duplicate post)

Thanks in advance!

[Sebastian Khan]

Hi Sebastian,

I see so what this is evaluate the inputs and ranks the features? But it does this in a greedy way right?

As a simple example suppose I have a 1D data set that is

$y(x) = 1 + x + x^3$

And I attempt to fit this with a basis of monomials I would expect the ideal set of bases to be $[x^0, x^1, x^3]$

Is this what SFS can do?

That is ultimately what I'm trying to do but then extend to up to a 7D case.

Thanks for your time!

Best,

Sebastian 

On Fri, 7 Feb 2020, 03:20 Sebastian Raschka, <se.ra...@me.com> wrote:

Hi,

So, the SFS will just select from an existing set of features. It will not perform any kind of feature transformation. You could use some arbitrary function though that manipulates your input data and add these as additional features to the set of existing ones.

Best,
Sebastian

> --
> You received this message because you are subscribed to the Google Groups "mlxtend" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to mlxtend+u...@googlegroups.com.
> To view this discussion on the web visit https://groups.google.com/d/msgid/mlxtend/b0f0f9d5-5c72-4bb3-a5dd-51814e1ae5aa%40googlegroups.com.