Add xgboost as engine for linear_reg()

[jaredlander]

{xgboost} can fit boosted, penalized linear models by setting booster="gblinear". This would be a great addition to linear_reg(), or perhaps there can be a boost_linear_reg() function and it can also include gbm().

Making this addition can probably take advantage of existing xgb.train() implementation because it only requires changing booster then using the lambda lambda_bias and alpha parameters consistently with the penalty and mixture arguments.

This then also brings up the question of fitting pseudo random forests with {xgboost}. Should this be added to rand_forest() or should the num_parallel_tree argument just be used when calling boost_tree()?

[martinocrippa]

Hi @juliasilge , as I wrote via e-mail, it could be useful to implement an xgboost model for regression with a linear booster. My target to predict derive from count or can be observed in a percentage way, I think it's better to use an ensemble model with a linear booster to test different distribution. My needs are to change the loss function because my target is skewness distributed, to manage it would be amazing if I will use linear booster parameters. thanks have a nice day MC

[ghost]

I would love to see a linear booster option added as well, thus enabling tuning for the three parameters listed with the linear booster in the xgboost package reference manual.