f_holdoutRF: Efficient Variable Importance Using Holdout Forests

In tim-coleman/RFtest: Scalable and Efficient Hypothesis Tests for Random Forests

Description

Runs variable importance for all variables by using a single forest. Traditional random forest works by restricting the features available to split upon in each node in the tree. Holdout forests modify this by instead restricting features available for the entire tree. This induces a natural sorting of variables into trees with and without the variables. Each collection of these trees forms a random forest, which is then run through the MSE Test procedure.

Usage

f_holdoutRF(X, y, X.test = FALSE, y.test = FALSE, B = 1000,
  NTest = nrow(X.test), mintree = 30, max.trees = 5 * ncol(X) * mintree,
  verbose = F, mtry = ncol(X)/3, p = 0.5, keep_forest = F, ...)

Arguments

X	Data frame of covariates - the training data.
y	Response vector. Currently only numeric responses (regression) are supported.
X.test	Covariates of the test set with which the MSE is calculated.
y.test	Responses in the test set with which the MSE is calculated.
base.learner	One of "rpart", "ctree", "rtree", or "lm". Base model to be used in the bagging.
single_forest	Logical. If TRUE, then all variables are compared against a single original forest.
NTest	If X.test, y.test are not specified, this number of test points are drawn at random from X, y.
Nbtree	How many trees should be used for each variable. Can either be a single number, or a vector of length ncol(X), with each entry corresponding to the number of trees used for testing each column.
verbose	Logical. Should a progress tracker be output in the console?
keep_forest	Logical. Should the original random forest be returned?
mtry	The mtry parameter used in random forest models.
p	Subsample size exponent, see MSE_Test.
...	Additional arguments to be passed to MSE_Test.
mintree
max.trees

tim-coleman/RFtest documentation built on March 10, 2020, 12:28 p.m.