Fix bug with Lrnr_grf predictions

[herbps10]

The .predict method for Lrnr_grf had an incorrect argument name for grf::predict.quantile_forest: the argument for supplying new data should be called newdata instead of new_data (see https://grf-labs.github.io/grf/reference/predict.quantile_forest.html).

[rachaelvp]

Thank you! This is great. Could you please add test to “tests/testthat/“ to make sure this is functioning as expected? Then I’ll merge PR

[nhejazi]

i'm confused by the diff including changes beyond the fix to Lrnr_grf but this is an important catch (thanks @herbps10!); could we double check changes and merge to devel @rachaelvp? happy to help with this

[herbps10]

Ah yes my mistake -- looks like I added an extra commit to the pull request by mistake, the only relevant one for this is 74f7538. Is there an easy way to fix this, or should I submit a new pull request? Thanks!

[nhejazi]

ok, thanks for clarifying @herbps10 -- depending on your level of comfort with git, you could rebase (e.g., https://stackoverflow.com/questions/36168839/how-to-remove-commits-from-a-pull-request), but opening up a new pull request is totally fine too

[herbps10]

Great, thanks for the tip -- I just fixed the PR. I can work on adding a unit test as well.

[nhejazi]

great, thanks! adding a unit test would be very helpful. i may periodically make some edits here and/or to the devel branch to get the automated builds restored before this is merged

[herbps10]

Quick question about how you would like Lrnr_grf to work:

With grf, if no new data are supplied for predictions (newdata = NULL), then out-of-bag predictions are returned for the training set (that is, predictions for the ith observation are made using only trees that did not use the ith training observation.)

As a consequence, if you supply the original training set as newdata (e.g. newdata = X.orig) then you get different predictions than if you predict with no new data (newdata = NULL), because the second one is out-of-bag and the first one is not.

The question is: when the predict method of Lrnr_grf is called with the same task as was used for training, should it return out-of-bag predictions or not?

(The current behavior, before the bugfix, is that the predict method of Lrnr_grf always returns the out-of-bag predictions for the training set, because newdata was always NULL).