Section 1: Data and model fits
Data used in this vignette:
The data used in the following examples is simulated from the Friedman benchmark problem 11. This benchmark problem is commonly used for testing purposes. The output is created according to the equation:
Create the data:
set.seed(101)
genFriedman <- function(noFeatures = 10,
noSamples = 100,
sigma = 1) {
# Set Values
n <- noSamples # no of rows
p <- noFeatures # no of variables
e <- rnorm(n, sd = sigma)
# Create matrix of values
xValues <- matrix(runif(n * p, 0, 1), nrow = n) # Create matrix
colnames(xValues) <- paste0("x", 1:p) # Name columns
df <- data.frame(xValues) # Create dataframe
# Equation:
# y = 10sin(πx1x2) + 20(x3−0.5)^2 + 10x4 + 5x5 + ε
y <- (10 * sin(pi * df$x1 * df$x2) + 20 * (df$x3 - 0.5)^2 + 10 * df$x4 + 5 * df$x5 + e)
# Adding y to df
df$y <- y
df
}
myData <- genFriedman(noFeatures = 9, noSamples = 350, sigma = 1)In the following examples, we use a ranger random forest model fit on the data, with the importance set to permutation.
set.seed(101)
fit <- randomForest(y ~ ., data = myData)Next, we create the ‘vivi-matrix’, which will contain variable importance on the diagonal and variable interactions in the upper and lower triangle. This matrix can then be supplied to the vivid plotting functions.
set.seed(101)
viFit <- vivi(
fit = fit,
data = myData,
response = "y",
gridSize = 10,
importanceType = NULL,
nmax = 100,
reorder = TRUE,
class = 1,
predictFun = NULL
)#Section 2: Visualizing the results
Heatmap plot
To create a heatmap of the vivi-matrix, we use:
viviHeatmap(mat = viFit) + ggtitle("random forest fit heatmap")
Network plot
To create a network graph of the vivi-matrix, we use:
viviNetwork(mat = viFit)
Generalized partial dependence pairs plot
In this plot we use a generalized pairs plot matrix style layout (which we call GPDP) to display partial dependence plots (PDPs) in the upper triangle, individual conditional exception curves (along with the aggregated 1-way partial dependence) on the diagonal and a scatterplot in the lower triangle.
To create the plot we supply the model fit to the plotting function.
set.seed(1701)
pdpPairs(data = myData, fit = fit, response = "y", nmax = 50, gridSize = 10)
#> Generating ice/pdp fits... waiting...
#> Finished ice/pdp
Partial dependence ‘Zenplot’
For this plot, we calculate the bivariate partial dependence and display them in a zenplots layout, which we call (ZPDP). The ZPDP is based on graph Eulerians and focuses on key subsets. ‘Zenplots’ create a zigzag expanded navigation plot (‘zenplot’) of the partial dependence values. This results in an alternating sequence of two-dimensional plots laid out in a zigzag structure, as shown in Fig 4.0 below and can be used as a useful space-saving plot that displays the most influential variables.
set.seed(1701)
pdpZen(data = myData, fit = fit, response = "y", nmax = 50, gridSize = 10)
#> Generating ice/pdp fits... waiting...
#> Finished ice/pdp
In Fig 4.0, we can see PDPs laid out in a zigzag structure, with the most influential variable pairs displayed at the top. As we move down the plot, we also move down in influence of the variable pairs.
Using the zpath argument, we can filter out any interactions below a set value. zpath takes the vivi matrix as a function argument and then, using cutoff, we can filter out any interactions below the chosen value. For example:
set.seed(1701)
zpath <- zPath(viv = viFit, cutoff = 0.1)
pdpZen(data = myData, fit = fit, response = "y", nmax = 50, gridSize = 10, zpath = zpath)
#> Generating ice/pdp fits... waiting...
#> Finished ice/pdp
