explain.tidyfit.models {tidyfit}

An interface for variable importance measures for a fitted tidyfit.models frames

Description

A generic method for calculating XAI and variable importance methods for tidyfit.models frames.

Usage

## S3 method for class 'tidyfit.models'
explain(
  object,
  use_package = NULL,
  use_method = NULL,
  ...,
  .keep_grid_id = FALSE
)

Arguments

object

model.frame created using regress, classify or m

use_package

the package to use to calculate variable importance. See 'Details' for possible options.

use_method

the method from 'use_package' that should be used to calculate variable importance.

...

additional arguments passed to the importance method

.keep_grid_id

boolean. By default the grid ID column is dropped, if there is only one unique setting per model or group. .keep_grid_id = TRUE ensures that the column is never dropped.

Details

WARNING This function is currently in an experimental stage.

The function uses the 'model_object' column in a tidyfit.model frame to return variable importance measures for each model.

Possible packages and methods include:

sensitivity package:

The package provides methods to assess variable importance in linear regressions ('lm') and classifications ('glm').

Usage: use_package="sensitivity" Methods:

• "lmg" (Shapley regression),

• "pmvd" (Proportional marginal variance decomposition),

• "src" (standardized regression coefficients),

• "pcc" (partial correlation coefficients),

• "johnson" (Johnson indices)

See ?sensitivity::lmg for more information and additional arguments.

iml package:

Integration with iml is currently in progress. The methods can be used for 'nnet', 'rf', 'lasso', 'enet', 'ridge', 'adalasso', 'glm' and 'lm'.

Usage: use_package="iml" Methods:

• "Shapley" (SHAP values)

• "LocalModel" (LIME)

• "FeatureImp" (Permutation-based feature importance)

The argument 'which_rows' (vector of integer indexes) can be used to explain specific rows in the data set for Shapley and LocalModel methods.

randomForest package:

This uses the native importance method of the randomForest package and can be used with 'rf' and 'quantile_rf' regression and classification.

Usage: use_package="randomForest" Methods:

• "mean_decrease_accuracy"

Value

A 'tibble'.

Author(s)

Johann Pfitzinger

References

Molnar C, Bischl B, Casalicchio G (2018). “iml: An R package for Interpretable Machine Learning.” JOSS, 3(26), 786. doi:10.21105/joss.00786.

Iooss B, Veiga SD, Janon A, Pujol G, Broto wcfB, Boumhaout K, Clouvel L, Delage T, Amri RE, Fruth J, Gilquin L, Guillaume J, Herin M, Idrissi MI, Le Gratiet L, Lemaitre P, Marrel A, Meynaoui A, Nelson BL, Monari F, Oomen R, Rakovec O, Ramos B, Rochet P, Roustant O, Sarazin G, Song E, Staum J, Sueur R, Touati T, Verges V, Weber F (2024). sensitivity: Global Sensitivity Analysis of Model Outputs and Importance Measures. R package version 1.30.0, https://CRAN.R-project.org/package=sensitivity.

A. Liaw and M. Wiener (2002). Classification and Regression by randomForest. R News 2(3), 18–22.

Examples

data <- dplyr::group_by(tidyfit::Factor_Industry_Returns, Industry)
fit <- regress(data, Return ~ ., m("lm"), .mask = "Date")
explain(fit, use_package = "sensitivity", use_method = "src")

data <- dplyr::filter(tidyfit::Factor_Industry_Returns, Industry == Industry[1])
fit <- regress(data, Return ~ ., m("lm"), .mask = c("Date", "Industry"))
explain(fit, use_package = "iml", use_method = "Shapley", which_rows = c(1))