PDPlot: Partial Dependence (PD) Plots

Description Usage Arguments Details Value Author(s) References See Also Examples

View source: R/PDPlot.R

Description

Computes and plots partial dependence (PD) plots for a fitted supervised learning model. The effects can be either a main effect for an individual predictor (length(J) = 1) or a second-order interaction effect for a pair of predictors (length(J) = 2).

Usage

1
PDPlot(X, X.model, pred.fun, J, K)

Arguments

X

The data frame of predictor variables to which the supervised learning model was fit. The names of the predictor variables must be the same as when the model was fit. The response variable should not be included in X.

X.model

The fitted supervised learning model object (e.g., a tree, random forest, neural network, etc.), typically an object to which a built-in predict command associated with that object can be applied.

pred.fun

A user-supplied function that will be used to predict the response for X.model for some specified inputs. pred.fun has two arguments. The first argument is named X.model and must be the same object as the X.model argument to the ALEPlot function. The second argument is named newdata and is a data frame of predictor values at which the object X.model is to be predicted. The output of pred.fun must be a numeric vector of predictions having length equal to the number of rows of newdata. For most X.model objects, pred.fun can simply call the predict function that was written as part of that modeling object package, assuming the package contains a predict function. An example of where a more customized pred.fun would be used is a multi (> 2) class classification problem for which the built-in predict function returns a vector of predicted probabilities, one for each response class. In this case it may make sense to have pred.fun return the predicted probabilities (or its log-odds, etc.) for one particular class of interest.

J

A numeric scalar or two-length vector of indices of the predictors for which the PD plot will be calculated. J is either a single index (for a main effects plot) or a pair of indices (for a second-order interaction plot). For a single index, the corresponding predictor must be either numeric or a factor. For a pair of indices, the corresponding predictors must be either both numeric or the first a factor and the second numeric.

K

A numeric scalar that represents the number of discrete points at which the PD plot will be calculated. If length(J) = 2, the same K will be used for both predictors, resulting in a two-dimensional grid of K^2 predictor values at with the PD plot will be calculated. K is only used if the predictor is numeric. For factor predictors, the equivalent of K is the number of used levels of the factor, which is automatically determined internally.

Details

This function calculates and plots the partial dependence (PD) plots first introduced in Friedman (2001). See the Apley (2016) reference paper listed below for details. For J = j (i.e., if the index for a single predictor x_j is specified), the function calculates and returns the PD main effect of x_j, which is denoted by f_j,PD(x_j) in Apley (2016). It also plots f_j,PD(x_j). For J = c(j1,j2) (i.e., if the indices for a pair of predictors (x_j1,x_j2) are specified), the function calculates and returns the PD second-order interaction effect of (x_j1,x_j2), which is denoted by f_{j1,j2},PD(x_j1,x_j2) in Apley (2016). It also plots f_{j1,j2},PD(x_j1,x_j2).

Value

f.values

If length(J) = 1, a vector of PD plot function values at the predictor values in x.values. If length(J) = 2, f.values is a K1xK matrix of the PD plot function values at the grid of predictor values defined by the X[,J[1]] and X[,J[2]] values in x.values. For X[,J[1]] numeric, K1 = K. For X[,J[2]] a factor, K1 is the number of used levels (empty levels are dropped). The rows of f.values correspond to X[,J(1)], and the columns to X[,J(2)].

x.values

For numeric predictors, if length(J) = 1, a K-length vector specifying the ordered predictor values at which the PD plot function is calculated. If length(J) = 2, a list of two such vectors, the first containing the X[,J(1)] values and the second containing the X[,J(2)] values at which the PD plot function is calculated. x.values is the same for factor predictors, except it is a K1-length character vector of the levels of the predictor, where K1 is determined internally as the number of unique levels of the predictor (empty levels are dropped).

Author(s)

Dan Apley

References

Friedman, J. H., (2001), "Greedy function approximation: A gradient boosting machine," Annals of Statistics, 29(5), pp. 1189-1232.

Apley, D. W. (2016), "Visualizing the Effects of Predictor Variables in Black Box Supervised Learning Models," submitted for publication.

See Also

See ALEPlot for partial dependence plots.

Examples

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
########################################################################
## A transparent example in which the supervised learning model is a linear regression \code{lm},
## but we will pretend it is black-box
########################################################################

## Generate some data and fit a \code{lm} supervised learning model
N=500
x1 <- runif(N, min=0, max=1)
x2 <- runif(N, min=0, max=1)
x3 <- runif(N, min=0, max=1)
y = x1 + 2*x2^2 + rnorm(N, 0, 0.1)
DAT = data.frame(y, x1, x2, x3)
lm.DAT = lm(y ~ .^2 + I(x1^2) + I(x2^2) + I(x3^2), DAT)

## Define the predictive function (easy in this case, since \code{lm} has 
## a built-in predict function that suffices)
yhat <- function(X.model, newdata) as.numeric(predict(X.model, newdata))

## Calculate and plot the PD main effects and second-order interaction effects of x1, x2, x3
par(mfrow = c(2,3))
PD.1=PDPlot(DAT[,2:4], lm.DAT, pred.fun=yhat, J=1, K=50)
PD.2=PDPlot(DAT[,2:4], lm.DAT, pred.fun=yhat, J=2, K=50)
PD.3=PDPlot(DAT[,2:4], lm.DAT, pred.fun=yhat, J=3, K=50)
PD.12=PDPlot(DAT[,2:4], lm.DAT, pred.fun=yhat, J=c(1,2), K=30)
PD.13=PDPlot(DAT[,2:4], lm.DAT, pred.fun=yhat, J=c(1,3), K=30)
PD.23=PDPlot(DAT[,2:4], lm.DAT, pred.fun=yhat, J=c(2,3), K=30)

## The following manually recreates the same plots produced by the above PDPlot function calls
par(mfrow = c(2,3))
plot(PD.1$x.values, PD.1$f.values, type="l", xlab="x1", ylab="PD main effect for x1")
plot(PD.2$x.values, PD.2$f.values, type="l", xlab="x2", ylab="PD main effect for x2")
plot(PD.3$x.values, PD.3$f.values, type="l", xlab="x3", ylab="PD main effect for x3")
image(PD.12$x.values[[1]], PD.12$x.values[[2]], PD.12$f.values, xlab = "x1", ylab = "x2")
contour(PD.12$x.values[[1]], PD.12$x.values[[2]], PD.12$f.values, add=TRUE, drawlabels=TRUE)
image(PD.13$x.values[[1]], PD.13$x.values[[2]], PD.13$f.values, xlab = "x1", ylab = "x3")
contour(PD.13$x.values[[1]], PD.13$x.values[[2]], PD.13$f.values, add=TRUE, drawlabels=TRUE)
image(PD.23$x.values[[1]], PD.23$x.values[[2]], PD.23$f.values, xlab = "x2", ylab = "x3")
contour(PD.23$x.values[[1]], PD.23$x.values[[2]], PD.23$f.values, add=TRUE, drawlabels=TRUE)


########################################################################
## A larger example in which the supervised learning model is a neural network (\code{nnet})
########################################################################

## Generate some data and fit a \code{nnet} supervised learning model

library(nnet)
N=5000
x1 <- runif(N, min=0, max=1)
x2 <- runif(N, min=0, max=1)
x3 <- runif(N, min=0, max=1)
y = x1 + 2*x2^2 +(x1-0.5)*(x3-0.5) + rnorm(N, 0, 0.1)
DAT = data.frame(y, x1, x2, x3)
nnet.DAT<-nnet(y~., data=DAT, linout=TRUE, skip=FALSE, size=10, decay=0.01, 
	maxit=1000, trace=FALSE)

## Define the predictive function
yhat <- function(X.model, newdata) as.numeric(predict(X.model, newdata, type="raw"))

## Calculate and plot the PD main and second-order interaction effects of x1, x2, x3
par(mfrow = c(2,3))
PD.1=PDPlot(DAT[,2:4], nnet.DAT, pred.fun=yhat, J=1, K=50)
PD.2=PDPlot(DAT[,2:4], nnet.DAT, pred.fun=yhat, J=2, K=50)
PD.3=PDPlot(DAT[,2:4], nnet.DAT, pred.fun=yhat, J=3, K=50)
PD.12=PDPlot(DAT[,2:4], nnet.DAT, pred.fun=yhat, J=c(1,2), K=20)
PD.13=PDPlot(DAT[,2:4], nnet.DAT, pred.fun=yhat, J=c(1,3), K=20)
PD.23=PDPlot(DAT[,2:4], nnet.DAT, pred.fun=yhat, J=c(2,3), K=20)
       
########################################################################
## A binary classification example in which the supervised learning model is 
## a neural network (\code{nnet}), and the log-odds of the predicted class 
## probability is the function to be plotted
########################################################################

## Generate some data and fit a \code{nnet} supervised learning model

library(nnet)
N=5000
x1 <- runif(N, min=0, max=1)
x2 <- runif(N, min=0, max=1)
x3 <- runif(N, min=0, max=1)
z = -3.21 + 2.81*x1 + 5.62*x2^2 + 2.81*(x1-0.5)*(x3-0.5) #true log-odds
p = exp(z)/(1+exp(z))
u = runif(N)
y = u < p
DAT = data.frame(y, x1, x2, x3)
nnet.DAT<-nnet(y~., data=DAT, linout=FALSE, skip=FALSE, size=10, decay=0.05, 
	maxit=1000, trace=FALSE)

## Define the ALE function to be the log-odds of the predicted probability that y = TRUE
yhat <- function(X.model, newdata) {
    p.hat = as.numeric(predict(X.model, newdata, type="raw"))
    log(p.hat/(1-p.hat))
}

## Calculate and plot the PD main and second-order interaction effects of x1, x2, x3
par(mfrow = c(2,3))
PD.1=PDPlot(DAT[,2:4], nnet.DAT, pred.fun=yhat, J=1, K=50)
PD.2=PDPlot(DAT[,2:4], nnet.DAT, pred.fun=yhat, J=2, K=50)
PD.3=PDPlot(DAT[,2:4], nnet.DAT, pred.fun=yhat, J=3, K=50)
PD.12=PDPlot(DAT[,2:4], nnet.DAT, pred.fun=yhat, J=c(1,2), K=20)
PD.13=PDPlot(DAT[,2:4], nnet.DAT, pred.fun=yhat, J=c(1,3), K=20)
PD.23=PDPlot(DAT[,2:4], nnet.DAT, pred.fun=yhat, J=c(2,3), K=20)

ALEPlot documentation built on May 1, 2019, 9:37 p.m.