modelStudio - perks and features

Hubert Baniecki

2023-08-31

Source: vignettes/ms-perks-features.Rmd

The modelStudio() function computes various (instance and dataset level) model explanations and produces a customisable dashboard, which consists of multiple panels for plots with their short descriptions. Easily save the dashboard and share it with others. Tools for Explanatory Model Analysis unite with tools for Exploratory Data Analysis to give a broad overview of the model behavior.

Let’s use HR dataset to explore modelStudio parameters:

train <- DALEX::HR
train$fired <- as.factor(ifelse(train$status == "fired", 1, 0))
train$status <- NULL

head(train)
DALEX::HR dataset
gender age hours evaluation salary fired
male 32.58 41.89 3 1 1
female 41.21 36.34 2 5 1
male 37.71 36.82 3 0 1
female 30.06 38.96 3 2 1
male 21.10 62.15 5 3 0
male 40.12 69.54 2 0 1

Prepare HR_test data and a ranger model for the explainer:

# fit a ranger model
library("ranger")
model <- ranger(fired ~., data = train, probability = TRUE)

# prepare validation dataset
test <- DALEX::HR_test[1:1000,]
test$fired <- ifelse(test$status == "fired", 1, 0)
test$status <- NULL

# create an explainer for the model
explainer <- DALEX::explain(model,
                            data = test,
                            y = test$fired)

# start modelStudio
library("modelStudio")

modelStudio parameters

instance explanations

Pass data points to the new_observation parameter for instance explanations such as Break Down, Shapley Values and Ceteris Paribus Profiles. Use new_observation_y to show their true labels.

new_observation <- test[1:3,]
rownames(new_observation) <- c("John Snow", "Arya Stark", "Samwell Tarly")
true_labels <- test[1:3,]$fired

modelStudio(explainer,
            new_observation = new_observation,
            new_observation_y  = true_labels)

If new_observation = NULL, then choose new_observation_n observations, evenly spread by the order of y_hat. This shall always include the observations, which ids are which.min(y_hat) and which.max(y_hat).

modelStudio(explainer, new_observation_n = 5) # default is 3

grid size

Achieve bigger or smaller modelStudio grid with facet_dim parameter.

# small dashboard with 2 panels
modelStudio(explainer,
            facet_dim = c(1,2))

# large dashboard with 9 panels
modelStudio(explainer,
            facet_dim = c(3,3))

animations

Manipulate time parameter to set animation length. Value 0 will make them invisible.

# slow down animations
modelStudio(explainer,
            time = 1000)

# turn off animations
modelStudio(explainer,
            time = 0)

more calculations means more time

Decrease N and B parameters to lower the computation time or increase them to get more accurate empirical results.

# faster, less precise
modelStudio(explainer,
            N = 200, B = 5)

# slower, more precise
modelStudio(explainer,
            N = 500, B = 15)

no EDA mode

Don’t compute the EDA plots if they are not needed. Set the eda parameter to FALSE.

modelStudio(explainer,
            eda = FALSE)

progress bar

Hide computation progress bar messages with show_info parameter.

modelStudio(explainer,
            show_info = FALSE)

viewer or browser?

Change viewer parameter to set where to display modelStudio. Best described in r2d3 documentation.

modelStudio(explainer,
            viewer = "browser")

parallel computation

Speed up modelStudio computation by setting parallel parameter to TRUE. It uses parallelMap package to calculate local explainers faster. It is really useful when using modelStudio with complicated models, vast datasets or many observations are being processed.

All options can be set outside of the function call. How to use parallelMap.

# set up the cluster
options(
  parallelMap.default.mode        = "socket",
  parallelMap.default.cpus        = 4,
  parallelMap.default.show.info   = FALSE
)

# calculations of local explanations will be distributed into 4 cores
modelStudio(explainer,
            new_observation = test[1:16,],
            parallel = TRUE)

additional options

Customize some of the modelStudio looks by overwriting default options returned by the ms_options() function. Full list of options.

# set additional graphical parameters
new_options <- ms_options(
  show_subtitle = TRUE,
  bd_subtitle = "Hello World",
  line_size = 5,
  point_size = 9,
  line_color = "pink",
  point_color = "purple",
  bd_positive_color = "yellow",
  bd_negative_color = "orange"
)

modelStudio(explainer,
            options = new_options)

All visual options can be changed after the calculations using ms_update_options().

old_ms <- modelStudio(explainer)
old_ms

# update the options
new_ms <- ms_update_options(old_ms,
                            time = 0,
                            facet_dim = c(1,2),
                            margin_left = 150)
new_ms

update observations

Use ms_update_observations() to add more observations with their local explanations to the modelStudio.

old_ms <- modelStudio(explainer)
old_ms

# add new observations
plus_ms <- ms_update_observations(old_ms,
                                  explainer,
                                  new_observation = test[101:102,])
plus_ms

# overwrite old observations
new_ms <- ms_update_observations(old_ms,
                                 explainer,
                                 new_observation = test[103:104,],
                                 overwrite = TRUE)
new_ms

Shiny

Use the widget_id argument and r2d3 package to render the modelStudio output in Shiny. See Using r2d3 with Shiny and consider the following example:

library(shiny)
library(r2d3)


ui <- fluidPage(
  textInput("text", h3("Text input"), 
            value = "Enter text..."),
  uiOutput('dashboard')
)

server <- function(input, output) {
  #:# id of div where modelStudio will appear
  WIDGET_ID = 'MODELSTUDIO'
  
  #:# create modelStudio 
  library(modelStudio)
  library(DALEX)
  model <- glm(survived ~., data = titanic_imputed, family = "binomial")
  explainer <- DALEX::explain(model,
                              data = titanic_imputed,
                              y = titanic_imputed$survived,
                              label = "Titanic GLM",
                              verbose = FALSE)
  ms <- modelStudio(explainer,
                    widget_id = WIDGET_ID,  #:# use the widget_id 
                    show_info = FALSE)    
  ms$elementId <- NULL                      #:# remove elementId to stop the warning

  #:# basic render d3 output
  output[[WIDGET_ID]] <- renderD3({
    ms
  })
  
  #:# use render ui to set proper width and height
  output$dashboard <- renderUI({
    d3Output(WIDGET_ID, width=ms$width, height=ms$height)
  })
}

shinyApp(ui = ui, server = server)

DALEXtra

Use explain_*() functions from the DALEXtra package to explain various models.

Bellow basic example of making modelStudio for a mlr model using explain_mlr().

library(DALEXtra)
library(mlr)

# fit a model
task <- makeClassifTask(id = "task", data = train, target = "fired")
learner <- makeLearner("classif.ranger", predict.type = "prob")
model <- train(learner, task)

# create an explainer for the model
explainer_mlr <- explain_mlr(model,
                             data = test,
                             y = test$fired,
                             label = "mlr")

# make a studio for the model
modelStudio(explainer_mlr)

References