In RifatMehreen/moccf: Counterfactual Fairness

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/README-",
  fig.height = 10,
  fig.width = 15
)
options(width = 200)

moccf: Multi-objective counterfactuals for counterfactual fairness

Installation

You can install the development version from GitHub with:

``` {r, results='hide', message=FALSE}

install.packages("devtools")

devtools::install_github("RifatMehreen/moccf")

## Getting started 

In this example, we train a `randomForest` on the `COMPAS` dataset.

We now examine whether the prediction of a given `yes` for `two_yr_recid` observation will change
to a `no` for the generated counterfactuals (with `race` changed to `Caucasian` from `African-American`).

``` {r, results='hide', message=FALSE}
library(dplyr)
library(plyr)
library(tidyverse)
library(Rtsne)
library(mlr3pipelines)
library(mlr3learners)
options(rgl.useNULL = TRUE)
library(rgl)
library(Rmpfr)
library(checkmate)
library(R6)
library(paradox)
library(data.table)
library(miesmuschel)
library(fairml)
library(counterfactuals)
library(randomForest)
library(iml)
library(ggforce)
library(moccf)

First, we load the data and pre-process it

compas <- fairml::compas
compas <- compas %>% drop_na()
compas <- compas %>% distinct()

Then we train the randomForest model to predict the two-yr-recid (Two years recidivism or chances of re-offending).
Note that we leave out one observation from the training data which is our x_interest.

set.seed(142)
rf = randomForest(two_year_recid ~ ., data = compas[-17L, ])

We now create an iml::Predictor object, that holds the model and the data for analyzing the model.

predictor = iml::Predictor$new(rf, type = "prob")

Now we set up an object of the FairnessTest that uses MOC from the counterfactuals package. As the sensitive attribute we use race.

fairness_obj = FairnessTest$new(predictor, df = compas, sensitive_attribute = "race", n_generations = 175)

For x_interest the model predicts:

x_interest = compas[17L, ]
predictor$predict(x_interest)

First we generate counterfactuals by running generate_countefactuals() for our x_interest. cfactuals is the genereted plausible counterfactuals. While we are examining unfairness for a specific protected attribute, we keep the other protected attributes fixed.

``` {r, results='hide', message=FALSE} cfactuals = fairness_obj$generate_counterfactuals(x_interest, desired_level = "Caucasian", desired_prob = c(0.5,1), fixed_features = "sex")

```r
cfactuals

We use the $get_prediction_difference() method to find differences of predictions of the x_interest and the cfactuals.

pred_diff = fairness_obj$get_prediction_difference(x_interest)
pred_diff

We can also use $get_mpd() method to look into the mean of the prediction differences

mpd = fairness_obj$get_mpd()
mpd

There are also some other methods for visualization and evaluation. In order to look into the distribution of the counterfactuals, we can use the $plot_tSNE() method. {r, message=FALSE} fairness_obj$plot_tSNE(x_interest, factor_variable = "sex")

RifatMehreen/moccf documentation built on Feb. 18, 2022, 1:35 a.m.