In dandls/counterfactuals: Counterfactual Explanations

In the following, we explain the counterfactuals workflow for both a classification and a regression task using concrete use cases.

# NOT_CRAN <- identical(tolower(Sys.getenv("NOT_CRAN")), "true")
knitr::opts_chunk$set(
  collapse = TRUE,
  fig.width = 7,
  fig.height = 3,
  comment = "#>"
  # purl = NOT_CRAN,
  # eval = NOT_CRAN
)
options(width = 200)

library("counterfactuals")
library("iml")
library("rpart")

knitr::opts_chunk$set(echo = TRUE)

Other types of models

The Predictor class of the iml package provides the necessary flexibility to cover classification and regression models fitted with diverse R packages. In the introduction vignette, we saw models fitted with the mlr3 and randomForest packages. In the following, we show extensions to - an classification tree fitted with the caret package, the mlr (a predecesor of mlr3) and tidymodels. For each model we generate counterfactuals for the 100th row of the plasma dataset of the gamlss.data package using the WhatIf method.

data(plasma, package = "gamlss.data")
x_interest = plasma[100L,]

rpart - caret package

library("caret")
treecaret = caret::train(retplasma ~ ., data = plasma[-100L,], method = "rpart", 
  tuneGrid = data.frame(cp = 0.01))
predcaret = Predictor$new(model = treecaret, data = plasma[-100L,], y = "retplasma")
predcaret$predict(x_interest)
nicecaret = NICERegr$new(predcaret, optimization = "proximity", 
  margin_correct = 0.5, return_multiple = FALSE)
nicecaret$find_counterfactuals(x_interest, desired_outcome = c(500, Inf))

rpart - tidymodels package

library("tidymodels")
treetm = decision_tree(mode = "regression", engine = "rpart") %>% 
  fit(retplasma ~ ., data = plasma[-100L,])
predtm = Predictor$new(model = treetm, data = plasma[-100L,], y = "retplasma")
predtm$predict(x_interest)
nicetm = NICERegr$new(predtm, optimization = "proximity", 
  margin_correct = 0.5, return_multiple = FALSE)
nicetm$find_counterfactuals(x_interest, desired_outcome = c(500, Inf))

rpart - mlr package

library("mlr")
task = mlr::makeRegrTask(data = plasma[-100L,], target = "retplasma")
mod = mlr::makeLearner("regr.rpart")

treemlr = mlr::train(mod, task)
predmlr = Predictor$new(model = treemlr, data = plasma[-100L,], y = "retplasma")
predmlr$predict(x_interest)
nicemlr = NICERegr$new(predmlr, optimization = "proximity", 
  margin_correct = 0.5, return_multiple = FALSE)
nicemlr$find_counterfactuals(x_interest, desired_outcome = c(500, Inf))

Decision tree - rpart package

treerpart = rpart(retplasma ~ ., data = plasma[-100L,])
predrpart = Predictor$new(model = treerpart, data = plasma[-100L,], y = "retplasma")
predrpart$predict(x_interest)
nicerpart = NICERegr$new(predrpart, optimization = "proximity", 
  margin_correct = 0.5, return_multiple = FALSE)
nicerpart$find_counterfactuals(x_interest, desired_outcome = c(500, Inf))