Joel Therrien
fdc708dad5
Add support for making predictions without specifying training data Add support for adding trees to an existing forest Add support for toggling displayProgress Also reduced the size of the package by removing some unused dependency classes.
183 lines
8.2 KiB
R
183 lines
8.2 KiB
R
% Generated by roxygen2: do not edit by hand
|
|
% Please edit documentation in R/train.R
|
|
\name{train}
|
|
\alias{train}
|
|
\alias{train.default}
|
|
\alias{train.formula}
|
|
\title{Train Random Forests}
|
|
\usage{
|
|
train(x, ...)
|
|
|
|
\method{train}{default}(responses, covariateData,
|
|
splitFinder = splitFinderDefault(responses),
|
|
nodeResponseCombiner = nodeResponseCombinerDefault(responses),
|
|
forestResponseCombiner = forestResponseCombinerDefault(responses),
|
|
ntree, numberOfSplits, mtry, nodeSize, maxNodeDepth = 1e+05,
|
|
splitPureNodes = TRUE, savePath = NULL,
|
|
savePath.overwrite = c("warn", "delete", "merge"),
|
|
cores = getCores(), randomSeed = NULL, displayProgress = TRUE)
|
|
|
|
\method{train}{formula}(formula, data, ...)
|
|
}
|
|
\arguments{
|
|
\item{responses}{An R list of the responses. See \code{\link{CR_Response}}
|
|
for an example function.}
|
|
|
|
\item{covariateData}{A data.frame containing only the columns of the
|
|
covariates you wish to use in your training (not relevant if you're using
|
|
the formula version of \code{train}).}
|
|
|
|
\item{splitFinder}{A split finder that's used to score splits in the random
|
|
forest training algorithm. See \code{\link{Competing Risk Split Finders}}
|
|
or \code{\link{WeightedVarianceSplitFinder}}. If you don't specify one,
|
|
this function tries to pick one based on the response. For
|
|
\code{\link{CR_Response}} without censor times, it will pick a
|
|
\code{\link{LogRankSplitFinder}}; while if censor times were provided it
|
|
will pick \code{\link{GrayLogRankSplitFinder}}; for integer or numeric
|
|
responses it picks a \code{\link{WeightedVarianceSplitFinder}}.}
|
|
|
|
\item{nodeResponseCombiner}{A response combiner that's used to combine
|
|
responses for each terminal node in a tree (regression example; average the
|
|
observations in each tree into a single number). See
|
|
\code{\link{CompetingRiskResponseCombiner}} or
|
|
\code{\link{MeanResponseCombiner}}. If you don't specify one, this function
|
|
tries to pick one based on the response. For \code{\link{CR_Response}} it
|
|
picks a \code{\link{CompetingRiskResponseCombiner}}; for integer or numeric
|
|
responses it picks a \code{\link{MeanResponseCombiner}}.}
|
|
|
|
\item{forestResponseCombiner}{A response combiner that's used to combine
|
|
predictions across trees into one final result (regression example; average
|
|
the prediction of each tree into a single number). See
|
|
\code{\link{CompetingRiskFunctionCombiner}} or
|
|
\code{\link{MeanResponseCombiner}}. If you don't specify one, this function
|
|
tries to pick one based on the response. For \code{\link{CR_Response}} it
|
|
picks a \code{\link{CompetingRiskFunctionCombiner}}; for integer or numeric
|
|
responses it picks a \code{\link{MeanResponseCombiner}}.}
|
|
|
|
\item{ntree}{An integer that specifies how many trees should be trained.}
|
|
|
|
\item{numberOfSplits}{A tuning parameter specifying how many random splits
|
|
should be tried for a covariate; a value of 0 means all splits will be
|
|
tried (with an exception for factors, who might have too many splits to
|
|
feasibly compute).}
|
|
|
|
\item{mtry}{A tuning parameter specifying how many covariates will be
|
|
randomly chosen to be tried in the splitting process. This value must be at
|
|
least 1.}
|
|
|
|
\item{nodeSize}{The algorithm will not attempt to split a node that has
|
|
observations less than 2*\code{nodeSize}; this guarantees that any two
|
|
sibling terminal nodes together have an average size of at least
|
|
\code{nodeSize}; note that it doesn't guarantee that every node is at least
|
|
as large as \code{nodeSize}.}
|
|
|
|
\item{maxNodeDepth}{This parameter is analogous to \code{nodeSize} in that it
|
|
controls tree length; by default \code{maxNodeDepth} is an extremely high
|
|
number and tree depth is controlled by \code{nodeSize}.}
|
|
|
|
\item{splitPureNodes}{This parameter determines whether the algorithm will
|
|
split a pure node. If set to FALSE, then before every split it will check
|
|
that every response is the same, and if so, not split. If set to TRUE it
|
|
forgoes that check and splits it. Prediction accuracy won't change under
|
|
any sensible \code{nodeResponseCombiner}; as all terminal nodes from a
|
|
split pure node should give the same prediction, so this parameter only
|
|
affects performance. If your response is continuous you'll likely
|
|
experience faster train times by setting it to TRUE. Default value is TRUE.}
|
|
|
|
\item{savePath}{If set, this parameter will save each tree of the random
|
|
forest in this directory as the forest is trained. Use this parameter if
|
|
you need to save memory while training. See also \code{\link{loadForest}}}
|
|
|
|
\item{savePath.overwrite}{This parameter controls the behaviour for what
|
|
happens if \code{savePath} is pointing to an existing directory. If set to
|
|
\code{warn} (default) then \code{train} refuses to proceed. If set to
|
|
\code{delete} then all the contents in that folder are deleted for the new
|
|
forest to be trained. Note that all contents are deleted, even those files
|
|
not related to \code{largeRCRF}. Use only if you're sure it's safe. If set
|
|
to \code{merge}, then the files describing the forest (such as its
|
|
parameters) are overwritten but the saved trees are not. The algorithm
|
|
assumes (without checking) that the existing trees are from a previous run
|
|
and starts from where it left off. This option is useful if recovering from
|
|
a crash.}
|
|
|
|
\item{cores}{This parameter specifies how many trees will be simultaneously
|
|
trained. By default the package attempts to detect how many cores you have
|
|
by using the \code{parallel} package and using all of them. You may specify
|
|
a lower number if you wish. It is not recommended to specify a number
|
|
greater than the number of available cores as this will hurt performance
|
|
with no available benefit.}
|
|
|
|
\item{randomSeed}{This parameter specifies a random seed if reproducible,
|
|
deterministic forests are desired.}
|
|
|
|
\item{displayProgress}{A logical indicating whether the progress should be
|
|
displayed to console; default is \code{TRUE}. Useful to set to FALSE in
|
|
some automated situations.}
|
|
|
|
\item{formula}{You may specify the response and covariates as a formula
|
|
instead; make sure the response in the formula is still properly
|
|
constructed; see \code{responses}}
|
|
|
|
\item{data}{A data.frame containing the columns of the predictors and
|
|
responses; not relevant if you're not using the formula version of
|
|
\code{train}.}
|
|
}
|
|
\value{
|
|
A \code{JRandomForest} object. You may call \code{predict} or
|
|
\code{print} on it.
|
|
}
|
|
\description{
|
|
Trains the random forest. The type of response the random forest can be
|
|
trained on varies depending on the \code{splitFinder},
|
|
\code{nodeResponseCombiner}, and the \code{forestResponseCombiner}
|
|
parameters. Make sure these are compatible with each other, and with the
|
|
response you plug in. \code{splitFinder} should work on the responses you are
|
|
providing; \code{nodeResponseCombiner} should combine these responses into
|
|
some intermediate product, and \code{forestResponseCombiner} combines these
|
|
intermediate products into the final output product. Note that
|
|
\code{nodeResponseCombiner} and \code{forestResponseCombiner} can be inferred
|
|
from the data (so feel free to not specify them), and \code{splitFinder} can
|
|
be inferred but you might want to change its default.
|
|
}
|
|
\note{
|
|
If saving memory is a concern, you can replace \code{covariateData} or
|
|
\code{data} with an environment containing one element called \code{data}
|
|
as the actual dataset. After the data has been imported into Java, but
|
|
before the forest training begins, the dataset in the environment is
|
|
deleted, freeing up memory in R.
|
|
}
|
|
\examples{
|
|
# Regression Example
|
|
x1 <- rnorm(1000)
|
|
x2 <- rnorm(1000)
|
|
y <- 1 + x1 + x2 + rnorm(1000)
|
|
|
|
data <- data.frame(x1, x2, y)
|
|
forest <- train(y ~ x1 + x2, data, WeightedVarianceSplitFinder(), MeanResponseCombiner(), MeanResponseCombiner(), ntree=100, numberOfSplits = 5, mtry = 1, nodeSize = 5)
|
|
|
|
# Fix x2 to be 0
|
|
newData <- data.frame(x1 = seq(from=-2, to=2, by=0.5), x2 = 0)
|
|
ypred <- predict(forest, newData)
|
|
|
|
plot(ypred ~ newData$x1, type="l")
|
|
|
|
# Competing Risk Example
|
|
x1 <- abs(rnorm(1000))
|
|
x2 <- abs(rnorm(1000))
|
|
|
|
T1 <- rexp(1000, rate=x1)
|
|
T2 <- rweibull(1000, shape=x1, scale=x2)
|
|
C <- rexp(1000)
|
|
u <- pmin(T1, T2, C)
|
|
delta <- ifelse(u==T1, 1, ifelse(u==T2, 2, 0))
|
|
|
|
data <- data.frame(x1, x2)
|
|
|
|
forest <- train(CR_Response(delta, u) ~ x1 + x2, data,
|
|
LogRankSplitFinder(1:2), CR_kResponseCombiner(1:2), CR_FunctionCombiner(1:2), ntree=100, numberOfSplits=5, mtry=1, nodeSize=10)
|
|
newData <- data.frame(x1 = c(-1, 0, 1), x2 = 0)
|
|
ypred <- predict(forest, newData)
|
|
}
|
|
\seealso{
|
|
\code{\link{predict.JRandomForest}}
|
|
}
|