Manually bagging ranger predictions

[mllg]

See https://stackoverflow.com/questions/65972567/how-can-i-save-all-prediction-from-ranger-learner-with-predit-all-true-into-a and https://github.com/mlr-org/mlr3learners/pull/172.

Should this be a pipe operator in mlr3pipelines? Maybe a nice use case to demonstrate how to write custom pipe ops?

[RaphaelS1]

I think it would be nice to see this as a much more general pipeline in which we have multiple predictions dependent on a set hyper-parameter, I've been discussing this with @adibender in the context of competing risks for proba.

So as a PipeOp in this learner we'd have something like:

• include the predict.all hyper-parameter
• add extra hyper-parameter to specify which tree is predict from
• use pipeop to iterate over all trees and return B predictions (for B trees)

For a competing risk PipeOp:

• include hyper-parameter for specifying competing risks
• add hyper-parameter to specify which event is predicted
• use pipe-op to iterate and return predictions for all events

So in the most general sense a PipeOp that iterates over all N values of a hyper-parameter to produce N prediction objects.