[R] Fix xgb.cv() for AFT models

[mayer79]

This PR fixes #7118

In order to keep the xgb.cv() API as it is:

• data must be an xgb.DMatrix object
• which contains the two infos 'label_lower_bound' and 'label_upper_bound'.

Automatic stratified splitting is deactivated with a warning. @david-cortes this is something we need to keep in mind for multioutput regressions as well.

[david-cortes]

Something I am wondering: most of these getinfo used throughout the function are only checking whether the DMatrix contains that field or not, but each call involves creating a full data copy just to check that it exists. @trivialfis Is there perhaps some function that could be used to just check if the DMatrix has the field or not? Do functions like XGDMatrixGetFloatInfo actually create a data copy when calling them?

[trivialfis]

Yes, it creates a data copy.

Hmm.. my personal preference is still to try not to use the DMatrix as a data manipulation class. Existing libraries and R builtins can do a much better job than we can ever try.

[david-cortes]

Thanks for looking into this. Left a small comment.

Although I am also thinking: if it constructs a DMatrix directly from data and labels - wouldn't it also require just a few lines to allow such a transformation with bounds as function arguments?

[mayer79]

We could add more arguments to xgb.cv(). I was actually thinking in the other direction: let the function work only with xgb.DMatrix input and remove the "label" argument. Why? The current signature of xgb.cv() is very incomplete. For instance there is no "weight" argument.

[david-cortes]

We could add more arguments to xgb.cv(). I was actually thinking in the other direction: let the function work only with xgb.DMatrix input and remove the "label" argument. Why? The current signature of xgb.cv() is very incomplete. For instance there is no "weight" argument.

Yes, that makes sense too - then we wouldn't need to update things in two places if new parameters come out.

[trivialfis]

cc @hcho3 for aft.

[david-cortes]

There's now a function xgb.DMatrix.hasinfo which could be used here to check whether the DMatrix has a given field without involving data copies:

xgboost/R-package/R/xgb.DMatrix.R

Line 223 in ff3d82c

xgb.DMatrix.hasinfo <- function(object, info) {

[mayer79]

Closed in favour of #10031