Does learning rate apply to linear booster?

[david-cortes]

There are some hints throughout the R examples for linear boosters, such as in the docs for xgb.cb.gblinear.history, which mention things like:

With the deterministic coordinate descent updater, it is safer to use higher learning rates.

xgboost/R-package/R/callbacks.R

Line 984 in 543b57f

#' # With the deterministic coordinate descent updater, it is safer to use higher learning rates.

The examples then pass learning_rate (previously eta) in the parameters:

xgboost/R-package/R/callbacks.R

Line 990 in 543b57f

#' learning_rate = 0.8,

.. but according to the docs about parameters, eta and learning_rate are for tree-based boosters, or at least that's how they are documented, since they aren't listed as parameters for the linear booster:

xgboost/doc/parameter.rst

Line 78 in 543b57f

* ``eta`` [default=0.3, alias: ``learning_rate``]

Does the linear booster use learning_rate / eta if passed under params?

[trivialfis]

Yes, it's used.

[trivialfis]

xgboost/src/linear/updater_coordinate.cc

Line 57 in 543b57f

auto dbias = static_cast<float>(tparam_.learning_rate *

[david-cortes]

Thanks. Sounds like the docs could be updated then, so as to list learning_rate under parameters for linear booster too. Other parameters like lambda are listed under both "tree booter" and "linear booster".

[trivialfis]

Thank you for updating the doc!

[david-cortes]

@trivialfis But I only updated the R docs. The online docs are still not showing the parameter under "linear booster". It's also not clear to me whether the default value is the same for both booster types either.