ranger {ranger}R Documentation

Ranger

Description

Ranger is a fast implementation of random forests (Breiman 2001) or recursive partitioning, particularly suited for high dimensional data. Classification, regression, and survival forests are supported. Classification and regression forests are implemented as in the original Random Forest (Breiman 2001), survival forests as in Random Survival Forests (Ishwaran et al. 2008). Includes implementations of extremely randomized trees (Geurts et al. 2006) and quantile regression forests (Meinshausen 2006).

Usage

ranger(
  formula = NULL,
  data = NULL,
  num.trees = 500,
  mtry = NULL,
  importance = "none",
  write.forest = TRUE,
  probability = FALSE,
  min.node.size = NULL,
  min.bucket = NULL,
  max.depth = NULL,
  replace = TRUE,
  sample.fraction = ifelse(replace, 1, 0.632),
  case.weights = NULL,
  class.weights = NULL,
  splitrule = NULL,
  num.random.splits = 1,
  alpha = 0.5,
  minprop = 0.1,
  split.select.weights = NULL,
  always.split.variables = NULL,
  respect.unordered.factors = NULL,
  scale.permutation.importance = FALSE,
  local.importance = FALSE,
  regularization.factor = 1,
  regularization.usedepth = FALSE,
  keep.inbag = FALSE,
  inbag = NULL,
  holdout = FALSE,
  quantreg = FALSE,
  time.interest = NULL,
  oob.error = TRUE,
  num.threads = NULL,
  save.memory = FALSE,
  verbose = TRUE,
  node.stats = FALSE,
  seed = NULL,
  dependent.variable.name = NULL,
  status.variable.name = NULL,
  classification = NULL,
  x = NULL,
  y = NULL,
  ...
)

Arguments

formula

Object of class formula or character describing the model to fit. Interaction terms supported only for numerical variables.

data

Training data of class data.frame, matrix, dgCMatrix (Matrix) or gwaa.data (GenABEL).

num.trees

Number of trees.

mtry

Number of variables to possibly split at in each node. Default is the (rounded down) square root of the number variables. Alternatively, a single argument function returning an integer, given the number of independent variables.

importance

Variable importance mode, one of 'none', 'impurity', 'impurity_corrected', 'permutation'. The 'impurity' measure is the Gini index for classification, the variance of the responses for regression and the sum of test statistics (see splitrule) for survival.

write.forest

Save ranger.forest object, required for prediction. Set to FALSE to reduce memory usage if no prediction intended.

probability

Grow a probability forest as in Malley et al. (2012).

min.node.size

Minimal node size to split at. Default 1 for classification, 5 for regression, 3 for survival, and 10 for probability.

min.bucket

Minimal terminal node size. No nodes smaller than this value can occur. Default 3 for survival and 1 for all other tree types.

max.depth

Maximal tree depth. A value of NULL or 0 (the default) corresponds to unlimited depth, 1 to tree stumps (1 split per tree).

replace

Sample with replacement.

sample.fraction

Fraction of observations to sample. Default is 1 for sampling with replacement and 0.632 for sampling without replacement. For classification, this can be a vector of class-specific values.

case.weights

Weights for sampling of training observations. Observations with larger weights will be selected with higher probability in the bootstrap (or subsampled) samples for the trees.

class.weights

Weights for the outcome classes (in order of the factor levels) in the splitting rule (cost sensitive learning). Classification and probability prediction only. For classification the weights are also applied in the majority vote in terminal nodes.

splitrule

Splitting rule. For classification and probability estimation "gini", "extratrees" or "hellinger" with default "gini". For regression "variance", "extratrees", "maxstat" or "beta" with default "variance". For survival "logrank", "extratrees", "C" or "maxstat" with default "logrank".

num.random.splits

For "extratrees" splitrule.: Number of random splits to consider for each candidate splitting variable.

alpha

For "maxstat" splitrule: Significance threshold to allow splitting.

minprop

For "maxstat" splitrule: Lower quantile of covariate distribution to be considered for splitting.

split.select.weights

Numeric vector with weights between 0 and 1, used to calculate the probability to select variables for splitting. Alternatively, a list of size num.trees, containing split select weight vectors for each tree can be used.

always.split.variables

Character vector with variable names to be always selected in addition to the mtry variables tried for splitting.

respect.unordered.factors

Handling of unordered factor covariates. One of 'ignore', 'order' and 'partition'. For the "extratrees" splitrule the default is "partition" for all other splitrules 'ignore'. Alternatively TRUE (='order') or FALSE (='ignore') can be used. See below for details.

scale.permutation.importance

Scale permutation importance by standard error as in (Breiman 2001). Only applicable if permutation variable importance mode selected.

local.importance

Calculate and return local importance values as in (Breiman 2001). Only applicable if importance is set to 'permutation'.

regularization.factor

Regularization factor (gain penalization), either a vector of length p or one value for all variables.

regularization.usedepth

Consider the depth in regularization.

keep.inbag

Save how often observations are in-bag in each tree.

inbag

Manually set observations per tree. List of size num.trees, containing inbag counts for each observation. Can be used for stratified sampling.

holdout

Hold-out mode. Hold-out all samples with case weight 0 and use these for variable importance and prediction error.

quantreg

Prepare quantile prediction as in quantile regression forests (Meinshausen 2006). Regression only. Set keep.inbag = TRUE to prepare out-of-bag quantile prediction.

time.interest

Time points of interest (survival only). Can be NULL (default, use all observed time points), a vector of time points or a single number to use as many time points (grid over observed time points).

oob.error

Compute OOB prediction error. Set to FALSE to save computation time, e.g. for large survival forests.

num.threads

Number of threads. Default is number of CPUs available.

save.memory

Use memory saving (but slower) splitting mode. No effect for survival and GWAS data. Warning: This option slows down the tree growing, use only if you encounter memory problems.

verbose

Show computation status and estimated runtime.

node.stats

Save node statistics. Set to TRUE to save prediction, number of observations and split statistics for each node.

seed

Random seed. Default is NULL, which generates the seed from R. Set to 0 to ignore the R seed.

dependent.variable.name

Name of dependent variable, needed if no formula given. For survival forests this is the time variable.

status.variable.name

Name of status variable, only applicable to survival data and needed if no formula given. Use 1 for event and 0 for censoring.

classification

Set to TRUE to grow a classification forest. Only needed if the data is a matrix or the response numeric.

x

Predictor data (independent variables), alternative interface to data with formula or dependent.variable.name.

y

Response vector (dependent variable), alternative interface to data with formula or dependent.variable.name. For survival use a Surv() object or a matrix with time and status.

...

Further arguments passed to or from other methods (currently ignored).

Details

The tree type is determined by the type of the dependent variable. For factors classification trees are grown, for numeric values regression trees and for survival objects survival trees. The Gini index is used as default splitting rule for classification. For regression, the estimated response variances or maximally selected rank statistics (Wright et al. 2016) can be used. For Survival the log-rank test, a C-index based splitting rule (Schmid et al. 2015) and maximally selected rank statistics (Wright et al. 2016) are available. For all tree types, forests of extremely randomized trees (Geurts et al. 2006) can be grown.

With the probability option and factor dependent variable a probability forest is grown. Here, the node impurity is used for splitting, as in classification forests. Predictions are class probabilities for each sample. In contrast to other implementations, each tree returns a probability estimate and these estimates are averaged for the forest probability estimate. For details see Malley et al. (2012).

Note that nodes with size smaller than min.node.size can occur because min.node.size is the minimal node size to split at, as in original Random Forests. To restrict the size of terminal nodes, set min.bucket. Variables selected with always.split.variables are tried additionally to the mtry variables randomly selected. In split.select.weights, weights do not need to sum up to 1, they will be normalized later. The weights are assigned to the variables in the order they appear in the formula or in the data if no formula is used. Names of the split.select.weights vector are ignored. Weights assigned by split.select.weights to variables in always.split.variables are ignored. The usage of split.select.weights can increase the computation times for large forests.

Unordered factor covariates can be handled in 3 different ways by using respect.unordered.factors: For 'ignore' all factors are regarded ordered, for 'partition' all possible 2-partitions are considered for splitting. For 'order' and 2-class classification the factor levels are ordered by their proportion falling in the second class, for regression by their mean response, as described in Hastie et al. (2009), chapter 9.2.4. For multiclass classification the factor levels are ordered by the first principal component of the weighted covariance matrix of the contingency table (Coppersmith et al. 1999), for survival by the median survival (or the largest available quantile if the median is not available). The use of 'order' is recommended, as it computationally fast and can handle an unlimited number of factor levels. Note that the factors are only reordered once and not again in each split.

The 'impurity_corrected' importance measure is unbiased in terms of the number of categories and category frequencies and is almost as fast as the standard impurity importance. It is a modified version of the method by Sandri & Zuccolotto (2008), which is faster and more memory efficient. See Nembrini et al. (2018) for details. This importance measure can be combined with the methods to estimate p-values in importance_pvalues. We recommend not to use the 'impurity_corrected' importance when making predictions since the feature permutation step might reduce predictive performance (a warning is raised when predicting on new data).

Note that ranger has different default values than other packages. For example, our default for mtry is the square root of the number of variables for all tree types, whereas other packages use different values for regression. Also, changing one hyperparameter does not change other hyperparameters (where possible). For example, splitrule="extratrees" uses randomized splitting but does not disable bagging as in Geurts et al. (2006). To disable bagging, use replace = FALSE, sample.fraction = 1. This can also be used to grow a single decision tree without bagging and feature subsetting: ranger(..., num.trees = 1, mtry = p, replace = FALSE, sample.fraction = 1), where p is the number of independent variables.

While random forests are known for their robustness, default hyperparameters not always work well. For example, for high dimensional data, increasing the mtry value and the number of trees num.trees is recommended. For more details and recommendations, see Probst et al. (2019). To find the best hyperparameters, consider hyperparameter tuning with the tuneRanger or mlr3 packages.

Out-of-bag prediction error is calculated as accuracy (proportion of misclassified observations) for classification, as Brier score for probability estimation, as mean squared error (MSE) for regression and as one minus Harrell's C-index for survival. Harrell's C-index is calculated based on the sum of the cumulative hazard function (CHF) over all timepoints, i.e., rowSums(chf), where chf is the the out-of-bag CHF; for details, see Ishwaran et al. (2008). Calculation of the out-of-bag prediction error can be turned off with oob.error = FALSE.

Regularization works by penalizing new variables by multiplying the splitting criterion by a factor, see Deng & Runger (2012) for details. If regularization.usedepth=TRUE, f^d is used, where f is the regularization factor and d the depth of the node. If regularization is used, multithreading is deactivated because all trees need access to the list of variables that are already included in the model.

For a large number of variables and data frames as input data the formula interface can be slow or impossible to use. Alternatively dependent.variable.name (and status.variable.name for survival) or x and y can be used. Use x and y with a matrix for x to avoid conversions and save memory. Consider setting save.memory = TRUE if you encounter memory problems for very large datasets, but be aware that this option slows down the tree growing.

For GWAS data consider combining ranger with the GenABEL package. See the Examples section below for a demonstration using Plink data. All SNPs in the GenABEL object will be used for splitting. To use only the SNPs without sex or other covariates from the phenotype file, use 0 on the right hand side of the formula. Note that missing values are treated as an extra category while splitting.

See https://github.com/imbs-hl/ranger for the development version.

With recent R versions, multithreading on Windows platforms should just work. If you compile yourself, the new RTools toolchain is required.

Value

Object of class ranger with elements

forest

Saved forest (If write.forest set to TRUE). Note that the variable IDs in the split.varIDs object do not necessarily represent the column number in R.

predictions

Predicted classes/values, based on out-of-bag samples (classification and regression only).

variable.importance

Variable importance for each independent variable.

variable.importance.local

Variable importance for each independent variable and each sample, if local.importance is set to TRUE and importance is set to 'permutation'.

prediction.error

Overall out-of-bag prediction error. For classification this is accuracy (proportion of misclassified observations), for probability estimation the Brier score, for regression the mean squared error and for survival one minus Harrell's C-index.

r.squared

R squared. Also called explained variance or coefficient of determination (regression only). Computed on out-of-bag data.

confusion.matrix

Contingency table for classes and predictions based on out-of-bag samples (classification only).

unique.death.times

Unique death times (survival only).

chf

Estimated cumulative hazard function for each sample (survival only).

survival

Estimated survival function for each sample (survival only).

call

Function call.

num.trees

Number of trees.

num.independent.variables

Number of independent variables.

mtry

Value of mtry used.

min.node.size

Value of minimal node size used.

treetype

Type of forest/tree. classification, regression or survival.

importance.mode

Importance mode used.

num.samples

Number of samples.

inbag.counts

Number of times the observations are in-bag in the trees.

dependent.variable.name

Name of the dependent variable. This is NULL when x/y interface is used.

status.variable.name

Name of the status variable (survival only). This is NULL when x/y interface is used.

Author(s)

Marvin N. Wright

References

See Also

predict.ranger

Examples

## Classification forest with default settings
ranger(Species ~ ., data = iris)

## Prediction
train.idx <- sample(nrow(iris), 2/3 * nrow(iris))
iris.train <- iris[train.idx, ]
iris.test <- iris[-train.idx, ]
rg.iris <- ranger(Species ~ ., data = iris.train)
pred.iris <- predict(rg.iris, data = iris.test)
table(iris.test$Species, pred.iris$predictions)

## Quantile regression forest
rf <- ranger(mpg ~ ., mtcars[1:26, ], quantreg = TRUE)
pred <- predict(rf, mtcars[27:32, ], type = "quantiles")
pred$predictions

## Variable importance
rg.iris <- ranger(Species ~ ., data = iris, importance = "impurity")
rg.iris$variable.importance

## Survival forest
require(survival)
rg.veteran <- ranger(Surv(time, status) ~ ., data = veteran)
plot(rg.veteran$unique.death.times, rg.veteran$survival[1,])

## Alternative interfaces (same results)
ranger(dependent.variable.name = "Species", data = iris)
ranger(y = iris[, 5], x = iris[, -5])

## Not run: 
## Use GenABEL interface to read Plink data into R and grow a classification forest
## The ped and map files are not included
library(GenABEL)
convert.snp.ped("data.ped", "data.map", "data.raw")
dat.gwaa <- load.gwaa.data("data.pheno", "data.raw")
phdata(dat.gwaa)$trait <- factor(phdata(dat.gwaa)$trait)
ranger(trait ~ ., data = dat.gwaa)

## End(Not run)


[Package ranger version 0.16.0 Index]