We need to figure out how we handle character features and levels

[mb706]

We should probably treat character features as something that may have an unlimited number of levels, leave it out from encoding, fixing factors etc. For factors and ordereds we should trust that levels(task$data()[[feature]]) is the same as task$levels()[[feature]]. In that case we can remove the "levels" argument from the data.table functions of PipeOpTaskPreproc.

[mllg]

We can do this, if this helps. Currently blocked by mlr3db: Many databases do not provide a native type for factors, everything is a character. At least we need an option there to auto-convert character -> factor in the backend.

[mb706]

Since mlr-org/mlr3#369 @mllg do you believe this is no longer blocked?

[mllg]

I guess so.

[mb706]

We now consistently handle character features as features without levels.