Hold out VIMP is calculated from the error rate for trees grown with and without a variable. Applies to all families.
# S3 method for rfsrc
holdout.vimp(formula, data,
ntree = function(p, vtry){1000 * p / vtry},
ntree.max = 2000,
ntree.allvars = NULL,
nsplit = 10,
ntime = 50,
mtry = NULL,
vtry = 1,
fast = FALSE,
verbose = TRUE,
...)A symbolic description of the model to be fit.
Data frame containing the y-outcome and x-variables.
Function specifying requested number of trees used for growing the forest. Inputs are dimension and number of holdout variables. The requested number of trees can also be a number.
Maximum number of trees used when calculating prediction error for determing hold out VIMP.
Grow this many additional trees and use them for
calculating the baseline error rate. Ignored if NULL.
Non-negative integer value specifying number of random split points used to split a node (deterministic splitting corresponds to the value zero and is much slower).
Integer value used for survival to
constrain ensemble calculations to a grid of ntime time points.
Number of variables randomly selected as candidates for splitting a node.
Number of variables randomly selected to be held out when growning a tree.
Use fast random forests, rfsrc.fast, in place of
rfsrc? Improves speed but is less accurate.
Provide verbose output?
Further arguments to be passed to rfsrc.
Hold out VIMP for each variable. For multivariate forests, hold out VIMP is calculated for each of the target outcomes.
Prior to growing a tree, a random set of vtry features are held
out. Tree growing proceeds as usual with the remaining features.
Once the forest is grown, hold out VIMP for a given variable v is
calculated as follows. Gather all trees where v was held out and
calculate OOB prediction error. Next gather all trees were v was not
held out and calculate OOB prediction error. Hold out VIMP for v is
the difference between these two values. Thus hold out VIMP measures
the importance of a variable when that variable is truly removed from
tree growing.
If ntree.allvars is set to an integer value, then a total of
this many trees are grown using all variables. The above procedure is
then implemented with the following change. Determine the error rate
for these additional trees. Hold out VIMP for v is the difference
between this value and the error rate for trees where v was held out.
Unlike the above procedure, this makes sure that the baseline used for
calculating holdout VIMP is the same for all v. This feature is
probably most useful in low-dimensional settings.
Note that accuracy of hold out VIMP depends heavily on the size of the
forest. If the number of trees is too small, then number of times a
variable is held out will be small and OOB error may suffer from high
variance. Thus, ntree should be set fairly high - we recommend
using 1000 times the number of features. Increasing vtry is
another way to increase number of hold out trees. In particular,
number of trees needed should decrease linearly with vtry. For
this reason the default ntree equals 1000 trees for each
feature divided by vtry. Keep in mind that intrepretation of
holdout VIMP is altered when vtry is different than 1.
Ishwaran H. (2019). Holdout variable importance for random forest models.
Lu M. and Ishwaran H. (2018). Expert Opinion: A prediction-based alternative to p-values in regression models. J. Thoracic and Cardiovascular Surgery, 155(3), 1130--1136.
# NOT RUN {
## ------------------------------------------------------------
## boston housing example
## ------------------------------------------------------------
if (library("mlbench", logical.return = TRUE)) {
data(BostonHousing)
hv <- holdout.vimp(medv ~ ., BostonHousing)
print(hv)
}
## ------------------------------------------------------------
## iris example illustrating vtry
## ------------------------------------------------------------
print(100 * holdout.vimp(Species ~ ., iris))
print(100 * holdout.vimp(Species ~ ., iris, vtry=2))
## ------------------------------------------------------------
## example involving class imbalanced data
## illustrates the new RFQ classifier
## see the function "imbalanced" for more information about RFQ
## ------------------------------------------------------------
data(breast, package = "randomForestSRC")
breast <- na.omit(breast)
f <- as.formula(status ~ .)
hv <- holdout.vimp(f, breast, rfq=TRUE, perf.type="g.mean")
print(10 * hv)
## ------------------------------------------------------------
## multivariate regression analysis example
## ------------------------------------------------------------
print(holdout.vimp(cbind(mpg, cyl) ~., mtcars))
## ------------------------------------------------------------
## white wine classification example
## ------------------------------------------------------------
data(wine, package = "randomForestSRC")
wine$quality <- factor(wine$quality)
hv <- holdout.vimp(quality ~ ., wine, vtry = 5)
print(100 * hv)
## ------------------------------------------------------------
## pbc survival example
## ------------------------------------------------------------
data(pbc, package = "randomForestSRC")
hv <- holdout.vimp(Surv(days, status) ~ ., pbc, splitrule = "random")
print(100 * hv)
## ------------------------------------------------------------
## WIHS competing risk example
## ------------------------------------------------------------
data(wihs, package = "randomForestSRC")
hv <- holdout.vimp(Surv(time, status) ~ ., wihs, ntree = 1000)
print(100 * hv)
# }
Run the code above in your browser using DataLab