modelStudio: Interactive Studio for Explanatory Model Analysis

Description

This 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.

The extensive documentation covers:

Function parameters description - perks and features
Framework and model compatibility - R & Python examples
Theoretical introduction to the plots - Explanatory Model Analysis: Explore, Explain, and Examine Predictive Models

Displayed variable can be changed by clicking on the bars of plots or with the first dropdown list, and observation can be changed with the second dropdown list. The dashboard gathers useful, but not sensitive, information about how it is being used (e.g. computation length, package version, dashboard dimensions). This is for the development purposes only and can be blocked by setting telemetry to FALSE.

Usage

modelStudio(explainer, ...)
# S3 method for explainer
modelStudio(
  explainer,
  new_observation = NULL,
  new_observation_y = NULL,
  new_observation_n = 3,
  facet_dim = c(2, 2),
  time = 500,
  max_features = 10,
  max_features_fi = NULL,
  N = 300,
  N_fi = N * 10,
  N_sv = N * 3,
  B = 10,
  B_fi = B,
  eda = TRUE,
  open_plots = c("fi"),
  show_info = TRUE,
  parallel = FALSE,
  options = ms_options(),
  viewer = "external",
  widget_id = NULL,
  license = NULL,
  telemetry = TRUE,
  max_vars = NULL,
  verbose = NULL,
  ...
)

Value

An object of the r2d3, htmlwidget, modelStudio class.

Arguments

explainer: An explainer created with DALEX::explain().
...: Other parameters.
new_observation: New observations with columns that correspond to variables used in the model.
new_observation_y: True label for new_observation (optional).
new_observation_n: Number of observations to be taken from the explainer$data if new_observation = NULL. See vignette
facet_dim: Dimensions of the grid. Default is c(2,2).
time: Time in ms. Set the animation length. Default is 500.
max_features: Maximum number of features to be included in BD, SV, and FI plots. Default is 10.
max_features_fi: Maximum number of features to be included in FI plot. Default is max_features.
N: Number of observations used for the calculation of PD and AD. Default is 300. See vignette
N_fi: Number of observations used for the calculation of FI. Default is 10*N.
N_sv: Number of observations used for the calculation of SV. Default is 3*N.
B: Number of permutation rounds used for calculation of SV. Default is 10. See vignette
B_fi: Number of permutation rounds used for calculation of FI. Default is B.
eda: Compute EDA plots and Residuals vs Feature plot, which adds the data to the dashboard. Default is TRUE.
open_plots: A vector listing plots to be initially opened (and on which positions). Default is c("fi").
show_info: Verbose a progress on the console. Default is TRUE.
parallel: Speed up the computation using parallelMap::parallelMap(). See vignette. This might interfere with showing progress using show_info.
options: Customize modelStudio. See ms_options and vignette.
viewer: Default is external to display in an external RStudio window. Use browser to display in an external browser or internal to use the RStudio internal viewer pane for output.
widget_id: Use an explicit element ID for the widget (rather than an automatically generated one). Useful e.g. when using modelStudio with Shiny. See vignette.
license: Path to the file containing the license (con parameter passed to readLines()). It can be used e.g. to include the license for explainer$data as a comment in the source of .html output file.
telemetry: The dashboard gathers useful, but not sensitive, information about how it is being used (e.g. computation length, package version, dashboard dimensions). This is for the development purposes only and can be blocked by setting telemetry to FALSE.
max_vars: An alias for max_features. If provided, it will override the value.
verbose: An alias for show_info. If provided, it will override the value.

References

The input object is implemented in DALEX
Feature Importance, Ceteris Paribus, Partial Dependence and Accumulated Dependence explanations are implemented in ingredients
Break Down and Shapley Values explanations are implemented in iBreakDown

Examples

Run this code

library("DALEX")
library("modelStudio")

#:# ex1 classification on 'titanic' data

# fit a model
model_titanic <- glm(survived ~., data = titanic_imputed, family = "binomial")

# create an explainer for the model
explainer_titanic <- explain(model_titanic,
                             data = titanic_imputed,
                             y = titanic_imputed$survived,
                             label = "Titanic GLM")

# pick observations
new_observations <- titanic_imputed[1:2,]
rownames(new_observations) <- c("Lucas","James")

# make a studio for the model
modelStudio(explainer_titanic,
            new_observations,
            N = 200,  B = 5) # faster example

# \donttest{

#:# ex2 regression on 'apartments' data
if (requireNamespace("ranger", quietly=TRUE)) {
  library("ranger")
  model_apartments <- ranger(m2.price ~. ,data = apartments)

  explainer_apartments <- explain(model_apartments,
                                  data = apartments,
                                  y = apartments$m2.price)

  new_apartments <- apartments[1:2,]
  rownames(new_apartments) <- c("ap1","ap2")

  # change dashboard dimensions and animation length
  modelStudio(explainer_apartments,
              new_apartments,
              facet_dim = c(2, 3),
              time = 800)

  # add information about true labels
  modelStudio(explainer_apartments,
              new_apartments,
              new_observation_y = new_apartments$m2.price)

  # don't compute EDA plots
  modelStudio(explainer_apartments,
              eda = FALSE)
}

#:# ex3 xgboost model on 'HR' dataset
if (requireNamespace("xgboost", quietly=TRUE)) {
  library("xgboost")
  HR_matrix <- model.matrix(status == "fired" ~ . -1, HR)

  # fit a model
  xgb_matrix <- xgb.DMatrix(HR_matrix, label = HR$status == "fired")
  params <- list(max_depth = 3, objective = "binary:logistic", eval_metric = "auc")
  model_HR <- xgb.train(params, xgb_matrix, nrounds = 300)

  # create an explainer for the model
  explainer_HR <- explain(model_HR,
                          data = HR_matrix,
                          y = HR$status == "fired",
                          type = "classification",
                          label = "xgboost")

  # pick observations
  new_observation <- HR_matrix[1:2, , drop=FALSE]
  rownames(new_observation) <- c("id1", "id2")

  # make a studio for the model
  modelStudio(explainer_HR,
              new_observation)
}
# }