OOBCurve (version 0.3)

OOBCurvePars: OOBCurvePars

Description

With the help of this function the out of bag curves for parameters like mtry, sample.fraction and min.node.size of random forests can be created for any measure that is available in the mlr package.

Usage

OOBCurvePars(lrn, task, pars = c("mtry"), nr.grid = 10, par.vals = NULL,
  measures = list(auc))

Arguments

lrn

The learner created with makeLearner. Currently only ranger is supported. num.trees has to be set sufficiently high to produce smooth curves.

task

Learning task created by the function makeClassifTask or makeRegrTask of mlr.

pars

One of the hyperparameter "mtry", "sample.fraction" or "min.node.size".

nr.grid

Number of points on hyperparameter space that should be evaluated (distributed equally)

par.vals

Optional vector of hyperparameter points that should be evaluated. If set, nr.grid is not used anymore. Default is NULL.

measures

List of performance measure(s) of mlr to evaluate. Default is mmce for classification and mse for regression. See the mlr tutorial for a list of available measures for the corresponding task.

Value

Returns a list with parameter values and a list of performances.

See Also

OOBCurve for out-of-bag curves dependent on the number of trees.

Examples

# NOT RUN {
library(mlr)
task = sonar.task

lrn = makeLearner("classif.ranger", predict.type = "prob", num.trees = 1000)
results = OOBCurvePars(lrn, task, measures = list(auc))
plot(results$par.vals, results$performances$auc, type = "l", xlab = "mtry", ylab = "auc")

lrn = makeLearner("classif.ranger", predict.type = "prob", num.trees = 1000, replace = FALSE)
results = OOBCurvePars(lrn, task, pars = "sample.fraction", measures = list(mmce))
plot(results$par.vals, results$performances$mmce, type = "l", xlab = "sample.fract.", ylab = "mmce")

results = OOBCurvePars(lrn, task, pars = "min.node.size", measures = list(mmce))
plot(results$par.vals, results$performances$mmce, type = "l", xlab = "min.node.size", ylab = "mmce")
# }

Run the code above in your browser using DataCamp Workspace