partial_dependence: Partial dependence using random forests

Description

Calculates the partial dependence of the response on an arbitrary dimensional set of predictors from a fitted random forest object from the party, randomForest, randomForestSRC, or ranger packages

Usage

partial_dependence(fit, vars, n, interaction, uniform, data, ...)

Arguments

fit

object of class 'RandomForest', 'randomForest', 'rfsrc', or `ranger`

vars

a character vector of the predictors of interest

two dimensional integer vector giving the resolution of the grid. the first element gives the grid on vars and the second on the other columns, which are subsampled.

interaction

logical, if 'vars' is a vector, does this specify an interaction or a list of bivariate partial dependence

uniform

logical, indicates whether a uniform or random grid is to be construct partial dependence calculation

data

the data.frame used to fit the model, only needed for 'randomForest'

...

additional arguments to be passed to marginalPrediction

Value

a data.frame with the partial dependence of 'vars' if 'vars' has length = 1 then the output will be a data.frame with a column for the predicted value at each value of 'vars', averaged over the values of all other predictors. if 'vars' has length > 1 and interaction is true or false then the output will be a data.frame with a column for each element of 'vars' and the predicted value for each combination.

References

Friedman, Jerome H. "Greedy function approximation: a gradient boosting machine." Annals of statistics (2001): 1189-1232.

Examples

Run this code

library(randomForest)
library(edarf)

data(iris)
data(swiss)

## classification
fit <- randomForest(Species ~ ., iris)
pd <- partial_dependence(fit, c("Sepal.Width", "Sepal.Length"),
  data = iris[, -ncol(iris)])
pd_int <- partial_dependence(fit, c("Petal.Width", "Sepal.Length"), 
  interaction = TRUE, data = iris[, -ncol(iris)])

## Regression
fit <- randomForest(Fertility ~ ., swiss)
pd <- partial_dependence(fit, c("Education", "Examination"), data = swiss[, -1])
pd_int <- partial_dependence(fit, c("Education", "Examination"),
  interaction = TRUE, data = swiss[, -1])