xgboost mtry parameter swap for #495

[topepo]

closes #495
closes #461

colsample_bytree remains an argument to xgb_train(). The mtry parameter in boost_tree() parameter now points to colsample_bynode.

We might want to add more engine-specific tunables to tune for this engine.

@mdancho84 I think that modeltime would need the same switch.

(edit for clarity)

[juliasilge]

So now both of these parameters are going to have the problem outlined in #461

[juliasilge]

I reran some analyses with xgboost, including this one, and the results are different in about the ways I think we'd expect:

No big changes in overall performance after tuning, but it switched some things in the variable importance and such. People who come back to train with the same data after this change are going to notice.

[mdancho84]

I'm here now - just seeing this. Yep, Modeltime will need the switch (I believe).

[mdancho84]

Hey, I've reviewed and I see one potential issue with backwards compatibility. What happens is that models that may have been specified with a value of 1 thinking this means 100%, gets converted to one column, which makes model performance very bad.

I recommend handling 1 as 100% of columns, and not 1 as 1 column.

The case in which a user actually only wants to use 1 columns should be rare, and handling as 100% is consistent with the underlying xgboost::xgb.train() function.

[topepo]

The previous behavior of "1.00 mean s 100% but otherwise it is a count" was a big mistake on my part.

It is 100% is consistent with xgboost, but completely inconsistent with randomForest, ranger, gbm, C50 and so on. We want to avoid people having to think "but for xgboost I have to do this instead of that". There are probably 4-5 things in the overall xgboost api that are irregular (and we fix them).

[mdancho84]

We are good to go. Let me know when parsnip is accepted by CRAN, and I will update modeltime.

[github-actions]

This pull request has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.