a36c3ed4f4
* Added SKLearn-like random forest Python API. - added XGBRFClassifier and XGBRFRegressor classes to SKL-like xgboost API - also added n_gpus and gpu_id parameters to SKL classes - added documentation describing how to use xgboost for random forests, as well as existing caveats
90 lines
3.7 KiB
ReStructuredText
90 lines
3.7 KiB
ReStructuredText
#########################
|
|
Random Forests in XGBoost
|
|
#########################
|
|
|
|
XGBoost is normally used to train gradient-boosted decision trees and other gradient
|
|
boosted models. Random forests use the same model representation and inference, as
|
|
gradient-boosted decision trees, but a different training algorithm. There are XGBoost
|
|
parameters that enable training a forest in a random forest fashion.
|
|
|
|
|
|
****************
|
|
With XGBoost API
|
|
****************
|
|
|
|
The following parameters must be set to enable random forest training.
|
|
|
|
* ``booster`` should be set to ``gbtree``, as we are training forests. Note that as this
|
|
is the default, this parameter needn't be set explicitly.
|
|
* ``subsample`` must be set to a value less than 1 to enable random selection of training
|
|
cases (rows).
|
|
* One of ``colsample_by*`` parameters must be set to a value less than 1 to enable random
|
|
selection of columns. Normally, ``colsample_bynode`` would be set to a value less than 1
|
|
to randomly sample columns at each tree split.
|
|
* ``num_parallel_tree`` should be set to the size of the forest being trained.
|
|
* ``num_boost_round`` should be set to 1. Note that this is a keyword argument to
|
|
``train()``, and is not part of the parameter dictionary.
|
|
* ``eta`` (alias: ``learning_rate``) must be set to 1 when training random forest
|
|
regression.
|
|
* ``random_state`` can be used to seed the random number generator.
|
|
|
|
|
|
Other parameters should be set in a similar way they are set for gradient boosting. For
|
|
instance, ``objective`` will typically be ``reg:linear`` for regression and
|
|
``binary:logistic`` for classification, ``lambda`` should be set according to a desired
|
|
regularization weight, etc.
|
|
|
|
If both ``num_parallel_tree`` and ``num_boost_round`` are greater than 1, training will
|
|
use a combination of random forest and gradient boosting strategy. It will perform
|
|
``num_boost_round`` rounds, boosting a random forest of ``num_parallel_tree`` trees at
|
|
each round. If early stopping is not enabled, the final model will consist of
|
|
``num_parallel_tree`` * ``num_boost_round`` trees.
|
|
|
|
Here is a sample parameter dictionary for training a random forest on a GPU using
|
|
xgboost::
|
|
|
|
params = {
|
|
'colsample_bynode': 0.8,
|
|
'learning_rate': 1,
|
|
'max_depth': 5,
|
|
'num_parallel_tree': 100,
|
|
'objective': 'binary:logistic',
|
|
'subsample': 0.8,
|
|
'tree_method': 'gpu_hist'
|
|
}
|
|
|
|
A random forest model can then be trained as follows::
|
|
|
|
bst = train(params, dmatrix, num_boost_round=1)
|
|
|
|
|
|
**************************
|
|
With Scikit-Learn-Like API
|
|
**************************
|
|
|
|
``XGBRFClassifier`` and ``XGBRFRegressor`` are SKL-like classes that provide random forest
|
|
functionality. They are basically versions of ``XGBClassifier`` and ``XGBRegressor`` that
|
|
train random forest instead of gradient boosting, and have default values and meaning of
|
|
some of the parameters adjusted accordingly. In particular:
|
|
|
|
* ``n_estimators`` specifies the size of the forest to be trained; it is converted to
|
|
``num_parallel_tree``, instead of the number of boosting rounds
|
|
* ``learning_rate`` is set to 1 by default
|
|
* ``colsample_bynode`` and ``subsample`` are set to 0.8 by default
|
|
* ``booster`` is always ``gbtree``
|
|
|
|
Note that these classes have a smaller selection of parameters compared to using
|
|
``train()``. In particular, it is impossible to combine random forests with gradient
|
|
boosting using this API.
|
|
|
|
|
|
*******
|
|
Caveats
|
|
*******
|
|
|
|
* XGBoost uses 2nd order approximation to the objective function. This can lead to results
|
|
that differ from a random forest implementation that uses the exact value of the
|
|
objective function.
|
|
* XGBoost does not perform replacement when subsampling training cases. Each training case
|
|
can occur in a subsampled set either 0 or 1 time.
|