Estimate Permutation p-values for Random Forest Importance Metrics
Estimate significance of importance metrics for a Random Forest model by permuting the response variable. Produces null distribution of importance metrics for each predictor variable and p-value of observed.
rfPermute(x, ...) ## Default S3 method: rfPermute(x, y, ..., nrep = 100, num.cores = 1) ## S3 method for class 'formula' rfPermute(formula, data = NULL, ..., subset, na.action = na.fail, nrep = 100)
x, y, formula, data, subset, na.action, ... |
See |
nrep |
Number of permutation replicates to run to construct null distribution and calculate p-values (default = 100). |
num.cores |
Number of CPUs to distribute permutation results over.
Defaults to |
All other parameters are as defined in randomForest.formula.
A Random Forest model is first created as normal to calculate the observed
values of variable importance. The response variable is then permuted
nrep times, with a new Random Forest model built for each
permutation step.
An rfPermute object which contains all of the components of a
randomForest object plus:
null.dist |
A list containing two three-dimensional arrays of null
distributions for |
pval |
A three dimensional array containing permutation p-values for
|
Eric Archer eric.archer@noaa.gov
plotNull for plotting null distributions from the rfPermute objects. rp.importance for extracting importance measures. rp.combine for combining multiple rfPermute objects.proximityPlot for plotting case proximities.impHeatmap for plotting a heatmap of importance scores.randomForest
# A regression model using the ozone example data(airquality) ozone.rfP <- rfPermute( Ozone ~ ., data = airquality, ntree = 100, na.action = na.omit, nrep = 50, num.cores = 1 ) # Plot the null distributions and observed values. plotNull(ozone.rfP) # Plot the unscaled importance distributions and highlight significant predictors plot(rp.importance(ozone.rfP, scale = FALSE)) # ... and the scaled measures plot(rp.importance(ozone.rfP, scale = TRUE))
Please choose more modern alternatives, such as Google Chrome or Mozilla Firefox.