Ranger

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,
  poisson.tau = 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,
  na.action = "na.learn",
  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. For classification, this can be a vector of class-specific values.
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. For classification, this can be a vector of class-specific values.
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", "beta" or "poisson" 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.
poisson.tau: For "poisson" splitrule: The coefficient of variation of the (expected) frequency is \(1/\tau\). If a terminal node has only 0 responses, the estimate is set to \(\alpha 0 + (1-\alpha) mean(parent)\) with \(\alpha = samples(child) mean(parent) / (\tau + samples(child) mean(parent))\).
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. Use 0 for all available cores. Default is 2 if not set by options/environment variables (see below).
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.
na.action: Handling of missing values. Set to "na.learn" to internally handle missing values (default, see below), to "na.omit" to omit observations with missing values and to "na.fail" to stop if missing values are found.
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).

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.

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.

Missing values can be internally handled by setting na.action = "na.learn" (default), by omitting observations with missing values with na.action = "na.omit" or by stopping if missing values are found with na.action = "na.fail". With na.action = "na.learn", missing values are ignored for calculating an initial split criterion value (i.e., decrease of impurity). Then for the best split, all missings are tried in both child nodes and the choice is made based again on the split criterion value. For prediction, this direction is saved as the "default" direction. If a missing occurs in prediction at a node where there is no default direction, it goes left.

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.

By default, ranger uses 2 threads. The default can be changed with: (1) num.threads in ranger/predict call, (2) environment variable R_RANGER_NUM_THREADS, (3) options(ranger.num.threads = N), (4) options(Ncpus = N), with precedence in that order.

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

References

Wright, M. N. & Ziegler, A. (2017). ranger: A fast implementation of random forests for high dimensional data in C++ and R. J Stat Softw 77:1-17. doi:10.18637/jss.v077.i01 .
Schmid, M., Wright, M. N. & Ziegler, A. (2016). On the use of Harrell's C for clinical risk prediction via random survival forests. Expert Syst Appl 63:450-459. doi:10.1016/j.eswa.2016.07.018 .
Wright, M. N., Dankowski, T. & Ziegler, A. (2017). Unbiased split variable selection for random survival forests using maximally selected rank statistics. Stat Med 36:1272-1284. doi:10.1002/sim.7212 .
Nembrini, S., Koenig, I. R. & Wright, M. N. (2018). The revival of the Gini Importance? Bioinformatics. doi:10.1093/bioinformatics/bty373 .
Breiman, L. (2001). Random forests. Mach Learn, 45:5-32. doi:10.1023/A:1010933404324 .
Ishwaran, H., Kogalur, U. B., Blackstone, E. H., & Lauer, M. S. (2008). Random survival forests. Ann Appl Stat 2:841-860. doi:10.1097/JTO.0b013e318233d835 .
Malley, J. D., Kruppa, J., Dasgupta, A., Malley, K. G., & Ziegler, A. (2012). Probability machines: consistent probability estimation using nonparametric learning machines. Methods Inf Med 51:74-81. doi:10.3414/ME00-01-0052 .
Hastie, T., Tibshirani, R., Friedman, J. (2009). The Elements of Statistical Learning. Springer, New York. 2nd edition.
Geurts, P., Ernst, D., Wehenkel, L. (2006). Extremely randomized trees. Mach Learn 63:3-42. doi:10.1007/s10994-006-6226-1 .
Meinshausen (2006). Quantile Regression Forests. J Mach Learn Res 7:983-999. https://www.jmlr.org/papers/v7/meinshausen06a.html.
Sandri, M. & Zuccolotto, P. (2008). A bias correction algorithm for the Gini variable importance measure in classification trees. J Comput Graph Stat, 17:611-628. doi:10.1198/106186008X344522 .
Coppersmith D., Hong S. J., Hosking J. R. (1999). Partitioning nominal attributes in decision trees. Data Min Knowl Discov 3:197-217. doi:10.1023/A:1009869804967 .
Deng & Runger (2012). Feature selection via regularized trees. The 2012 International Joint Conference on Neural Networks (IJCNN), Brisbane, Australia. doi:10.1109/IJCNN.2012.6252640 .
Probst, P., Wright, M. N. & Boulesteix, A-L. (2019). Hyperparameters and tuning strategies for random forest. WIREs Data Mining Knowl Discov 9:e1301.doi:10.1002/widm.1301 .

Author

Marvin N. Wright

Examples

## Classification forest with default settings
ranger(Species ~ ., data = iris)
#> Ranger result
#> 
#> Call:
#>  ranger(Species ~ ., data = iris) 
#> 
#> Type:                             Classification 
#> Number of trees:                  500 
#> Sample size:                      150 
#> Number of independent variables:  4 
#> Mtry:                             2 
#> Target node size:                 1 
#> Variable importance mode:         none 
#> Splitrule:                        gini 
#> OOB prediction error:             5.33 % 

## 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)
#>             
#>              setosa versicolor virginica
#>   setosa         13          0         0
#>   versicolor      0         17         1
#>   virginica       0          0        19

## Quantile regression forest
rf <- ranger(mpg ~ ., mtcars[1:26, ], quantreg = TRUE)
pred <- predict(rf, mtcars[27:32, ], type = "quantiles")
pred$predictions
#>      quantile= 0.1 quantile= 0.5 quantile= 0.9
#> [1,]         21.00          22.8         32.55
#> [2,]         21.00          22.8         30.40
#> [3,]         13.30          15.5         30.40
#> [4,]         15.47          21.0         22.80
#> [5,]         13.30          14.3         21.00
#> [6,]         21.00          22.8         32.40

## Variable importance
rg.iris <- ranger(Species ~ ., data = iris, importance = "impurity")
rg.iris$variable.importance
#> Sepal.Length  Sepal.Width Petal.Length  Petal.Width 
#>     9.818318     2.371400    44.752239    42.281016 

## Survival forest
require(survival)
#> Loading required package: 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 result
#> 
#> Call:
#>  ranger(dependent.variable.name = "Species", data = iris) 
#> 
#> Type:                             Classification 
#> Number of trees:                  500 
#> Sample size:                      150 
#> Number of independent variables:  4 
#> Mtry:                             2 
#> Target node size:                 1 
#> Variable importance mode:         none 
#> Splitrule:                        gini 
#> OOB prediction error:             4.67 % 
ranger(y = iris[, 5], x = iris[, -5])
#> Ranger result
#> 
#> Call:
#>  ranger(y = iris[, 5], x = iris[, -5]) 
#> 
#> Type:                             Classification 
#> Number of trees:                  500 
#> Sample size:                      150 
#> Number of independent variables:  4 
#> Mtry:                             2 
#> Target node size:                 1 
#> Variable importance mode:         none 
#> Splitrule:                        gini 
#> OOB prediction error:             4.67 % 

if (FALSE) { # \dontrun{
## 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)
} # }