In AlanInglis/vividOld: Variable Importance and Variable Interaction Displays

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.width = 7,
  fig.height = 7
)

This vignette is intended to be used as a supplementary tools to the introduction to the vivid package vignette. The goal of this vignette is to show how vivid handles a range of different response variables and to show a brief guide on how to create visualisations from the package to display varible importance and variable interactions (VIVI). To begin we show a standard response from a random forest regression model.

First, we load the required packages. vivid to create the visualizations and mlr3 and mlr3learners will be used to create our models.

library(vivid)
library(mlr3)  
library(mlr3learners)

Regression with a continuous response

For this we will simulate some data from the Friedman benchmark problem 1 using the genFriedman() function from vivid, with 9 features and 300 observations. The output is created according to:

$$y = 10 sin(π x1 x2) + 20 (x3 - 0.5)^2 + 10 x4 + 5 x5 + e$$

Create the data:

set.seed(1701)
myData <- genFriedman(noFeatures = 9, noSamples = 300, sigma = 1, bins = NULL, seed = NULL)

Create an mlr3 ranger random forest model:

set.seed(1701)
taskCont  <- TaskRegr$new(id = "Friedman", backend = myData, target = "y")
set.seed(1701)
lrnCont <- lrn("regr.ranger", importance = "permutation")
set.seed(1701)
modCont <- lrnCont$train(taskCont)

Create a matrix to be supplied to the plotting functions.

set.seed(1701)
myMatrix <- vividMatrix(task = taskCont, model = modCont, gridSize = 100)

Create a heatmap of the results

plot(myMatrix, type = "heatMap")

Fig 1.0: Heat-map plot displaying 2-way interaction strength in blue and individual variable importance on the diagonal in red.

Figure 1.0 shows a heatmap of VIVI values on the Friedman data. We can clearly see the variables $x_1$ to $x_5$ are the only important variables for predeicting $y$, with $x_4$ being the most important. We can also seee a strong interaction between $x_1$ and $x_2$ Create a partial dependence pairs plot (GPDP)

ggpdpPairs(taskCont, modCont, gridsize = 20)

Create a zenplot partial dependence plot (ZPDP)

pdpZenplot(taskCont, modCont, gridsize = 20)

Numerical Binary Response

Here we can use genFriedman again to create data with a binary response. To do this we set the number of bins to be equal to 2. This will create (roughly) equal sized bins to split the response into. Setting a value greater than 1 essentially turns this into a classification problem where bins determines the number of classes.

Create the data:

set.seed(1701)
myDataBin <- genFriedman(noFeatures = 9, noSamples = 300, sigma = 1, bins = 2, seed = NULL)
myDataBin$y <- as.numeric(myDataBin$y)

Take a quick look at our data to see that the response has been split into 2 separate classes.

head(myDataBin)

Create an mlr3 ranger random forest model:

set.seed(1701)
taskBin  <- TaskRegr$new(id = "Friedman", backend = myDataBin, target = "y")
set.seed(1701)
lrnBin <- lrn("regr.ranger", importance = "impurity")
set.seed(1701)
modBin <- lrnCont$train(taskCont)

Create a matrix to be supplied to the plotting functions.

set.seed(1701)
myMatrixBin <- vividMatrix(task = taskBin, model = modBin, gridSize = 100)

Create a heatmap of the results

plot(myMatrixBin, type = "heatMap")

Create a partial dependence pairs plot (PDPP)

ggpdpPairs(taskBin, modBin, gridsize = 20)

Create a zenplot partial dependence plot (ZPDP)

pdpZenplot(taskBin, modBin, gridsize = 20)

Multiclass Categorical Response

For this we will use the iris dataset to show how to use vivd when there is a multiclass categorical response.

To begin, we create our mlr3 model.

# mlr3 TASK
set.seed(1701)
ir_T <- TaskClassif$new(id = "iris", backend = iris, target = "Species")
# learner
set.seed(1701)
lrn <- lrn("classif.ranger", importance = 'impurity')
# model
set.seed(1701)
ir_M <- lrn$train(ir_T)

To create our vivid matrix of values, we have to specify the main variable of interest by use of the main argument. In the example below, we choose to look at the species "virginica".

myMatrixMCR <- vividMatrix(ir_T, ir_M, gridsize = 100,  main = "virginica")

Plot the results:

plot(myMatrixMCR, type = 'heatMap', title = "virginica")

Create a partial dependence pairs plot (PDPP)

ggpdpPairs(ir_T, ir_M, gridsize = 20)

Create a zenplot partial dependence plot (ZPDP)

pdpZenplot(ir_T, ir_M, gridsize = 20)

AlanInglis/vividOld documentation built on March 4, 2021, 12:44 a.m.