search

mlr-org / mlr3filters

Filter-based feature selection for mlr3
https://mlr3filters.mlr-org.com
GNU Lesser General Public License v3.0
20 stars 8 forks source link

Think about how to implement PCA #31

Closed pat-s closed 5 years ago

pat-s commented 5 years ago

In mlr we used makePreprocWrapperCaret().

mllg commented 5 years ago

There was a filter for PCA in mlr?

pat-s commented 5 years ago

The only reference I could find was via makePreprocWrapperCaret().

We may want to use a package for that instead of doing the work on our own.

mllg commented 5 years ago

I don't get what this filter is supposed to do. Do you want to extract features based on their loadings on the PCs?

You can do PCA with prcomp() or princomp() from stats.

jakob-r commented 5 years ago

Just to rule out the obvious: We are not talking about the normal PCA that is a PipeOp right?

berndbischl commented 5 years ago

I also don't get what a pca filter should do. If no one can link to a rather standard paper that discusses this, the issue should be closed imho

mllg commented 5 years ago

Well, the score could be the explained variance. That would not be too extraordinary.

pat-s commented 5 years ago

I guess my initial thinking when opening this issue was that all operations that perform "feature selection" in any way should be placed in here - when the package was still called "mlr3featsel" and we wanted to include wrappers, filters and PCA (on the same level).

Now lots has changed and we only have filters in here and yes, PCA is already supported via a PipeOp.

I did not have in mind implementing PCA as a filter (I also do not see how that should make sense) but as a standalone "feature selection" method.

berndbischl commented 5 years ago

Michel said. Well, the score could be the explained variance. That would not be too extraordinary.

Well that does make sense. If someone wants this, this could be implemented