RFinfer (version 0.2.0)

rfPredVar: rfPredVar

Description

Generate predictions and prediction variances from a random forest based on the infinitesimal jackknife.

Usage

rfPredVar(random.forest, rf.data, pred.data = rf.data, CI = FALSE, tree.type = "rf", prog.bar = FALSE)

Arguments

random.forest

A random forest trained with keep.inbag=TRUE. See details for more information.

rf.data

The data used to train rf

pred.data

The data used to predict with the forest; defaults to rf.data if not given

CI

Should 95% confidence intervals based on the CLT be returned along with predictions and prediction variances?

tree.type

either 'ci' for conditional inference tree or 'rf' for traditional CART tree

prog.bar

should progress bar be shown? (only applicable when tree.type='ci')

Value

A data frame with the predictions and prediction variances (and optionally 95% confidence interval)

Details

The random forest trained with keep.inbag=TRUE is supplied only for the purpose of defining the resampling scheme. The function builds a new random forest based on the tree.type setting. However, the resamples are maintained identically to the supplied random forest. This allows for direct comparison of the tree methods without having to account for variation in resampling.

Currently, the CI methods are much more computationally intensive because there is no C implementation of the CI random forest method that indicates the number of times that each sample is included in each resample. In order to carry out our simulations using $V_IJ^B$, we had to use a pure R implementation of CI random forests. This is different for CART random forests, where a C implementation already exists in the randomForest package. However, it should be noted that the difference in computational times is due to the random forest creation step, not the implementation of $V_IJ^B$. This should not be an issue in the future when a C implementation of CI random forests is created.

Note: This function does not use the default predict method for forests produced by cforest. The predictions here are the direct averages of all tree predictions, instead of using the observation weights. Therefore, predictions from this function will likely differ from predict.cforest when using subsampling.

This function currently only works with regression forests -- not classification forests.

Examples

library(randomForest)
data(airquality)
d <- na.omit(airquality)
rf <- randomForest(Ozone ~ .,data=d,keep.inbag=TRUE,sampsize=30,replace=FALSE,ntree=500)
rfPredVar(rf,rf.data=d,CI=TRUE,tree.type='rf')