R package produces NULL getinfo(dtrain, "group")

[andrefontenele]

While performing cross validation using xgb.cv with a custom_metric, getinfo(dtrain, "group") always returns NULL making it impossible to identify which folder the algorithm is currently using. I have tried this in numerous setups.

[trivialfis]

If I remember correctly, there's no group meta info. The similar one group_ptr stores the query group for ranking tasks. Could you please elaborate on what you are trying to achieve.

[andrefontenele]

Thank you so much for the reply. Ah, i see, so group has nothing to see with the fold.

I am trying to calculate a performance metric that depends on some information that is outside the model per se (ie, something that is not in the DMatrix). It is basically a different version of the label (numeric instead of categorical). Is that possible? No information that I can get from getinfo() (ie 'label', 'weight', 'base_margin', 'label_lower_bound', 'label_upper_bound', 'group', 'feature_type', 'feature_name', 'nrow') seems to allow me to identify which fold the cross validation is currently working on, not even if it is in the training or the testing part.

[trivialfis]

XGBoost computes all folds for each new boosting iteration:

xgboost/R-package/R/xgb.cv.R

Line 249 in 850d15a

msg <- lapply(bst_folds, function(fd) {

If you want to have custom evaluation, you can try the callback function, which accepts all folds as input:

xgboost/R-package/R/xgb.cv.R

Line 265 in 850d15a

should_stop <- .execute.cb.after.iter(

[andrefontenele]

Ok, thank you very much for the reply. Apologies for not understanding the issue properly, but is was really hard to grasp it directly from the documentation and LLMs.

[trivialfis]

No worries. If you have specific questions, feel free to ask. Or, better yet, suggestions for documentation are welcome.