partialDep: Partial Dependence (single variable)

In brooksandrew/Rsenal: Rsenal (arsenal) of assorted R functions developed over the years

Description

Calculate the partial dependence of a predictor variable on the response variable from a random forest classification model. Rather than sequence through values of the predictor variable of interest and keep the other predictors at their median, this partial dependence technique creates replicates of the entire dataset for each level of the x variable of interest from it's min to max. This gives a more realistic idea of the magnitude and direction of the x variable on the response.

Usage

partialDep(model, df, xvar, n = 10, target.class = "1", ci = c(0.9, 0.5,
  0.3))

Arguments

model	model object used to generate predictions. Currently only built and tested for random forest.
df	data.frame or data.table used to generate predictions with model
xvar	character of length one; the x variable in df to assess for partial dependence with the response variable from model.
n	numeric of length one; number of values between the min and max of xvar to score the model. Note: this number is also how many replicates of df must be created and stored in memory. Default is 10.
target.class	character: Which category (class) of the target variable to use for predictions
ci	numeric: specify any confidence intervals around the median response.

Value

data.table of output. cnt refers to how many obs from df are within the fixed-width bin specified by xvar.

See Also

partialDepPlot

Examples

library('randomForest')
library('data.table')
DF <- mtcars
DF$vs <- factor(DF$vs)
rf <- randomForest(vs~mpg+cyl+drat+qsec+disp+gear+carb+hp, DF, ntrees=100)
pd <- partialDep(model=rf, df=DF, xvar='mpg')
pd[ci==0.5,] # median of response when sequenced through 'mpg'

## Plotting
plot(pd[cilev==0, xvar], pd[cilev==0, pred], type='l', ylim=c(0,1))
lines(pd[ci==.95, xvar], pd[ci==.95, pred], type='l', col='red')
lines(pd[ci==.05, xvar], pd[ci==.05, pred], type='l', col='green')

brooksandrew/Rsenal documentation built on May 13, 2019, 7:50 a.m.