Gblinear parameters

#1

[migrated from https://github.com/dmlc/xgboost/issues/3492]

Few questions about gblinear:

  • Is there a parameter for feature subsetting like “colsample_bytree” for gblinear?
  • Is gblinear with default parameters functioning like an elastic net?
  • What are the range of values for alpha and lambda, they are not listed in https://xgboost.readthedocs.io/en/latest/parameter.html.
    Thanks!
#2

No, currently there is no equivalent feature for gblinear.

No, since the default values for both reg_lambda (L2 regularization) and reg_alpha (L1 regularization) default to zero. For elastic nets, set both values to nonzero.

Any non-negative values will do.

#3

I there any way to calculate a non negative least square regression with XGBoost?