Feature request: Support mixed model feature extraction

[adibender]

Goal is to have random effect estimates of a mixed model as features in the model

Should be relatively straight forward

1. Transform data to long format (i guess there is something in tf package)
2. Fit Mixed model with where the functional feature is the outcome and id the input using nlme or lme4 (formula would need to be supplied, but usually we want random intercept and random slope)
3. "predict" the random effects (intercept + slope) for each id
4. return

[m-muecke]

You can directly unnest with as.data.frame.tf with unnest set to true.

I would prefer the use of lme4, since Fabian is also a contributor of it.

So I'm guessing you want something like the following: value ~ arg + (arg | id), but make the formula a flexible user input?

[adibender]

lme4 would be fine, but I think within the pipeline one could use different mlr3learners that are available for mixed modeling, so we don't need to restrict ourselves?
@markusgoeswein is currently working on this

[m-muecke]

lme4 would be fine, but I think within the pipeline one could use different mlr3learners that are available for mixed modeling, so we don't need to restrict ourselves? @markusgoeswein is currently working on this

Technically, sure. But, there is currently no learner property for mixed modeling. So either the learner options are checked for hardcoded values or the user just provides the backend option such as lme4 or others.