Fair Data Adaptation with Quantile Preservation

An implementation of the fair data adaptation with quantile preservation described in Plecko & Meinshausen (JMLR 2020, 21(242), 1-44). The adaptation procedure uses the specified causal graph to pre-process the given training and testing data in such a way to remove the bias caused by the protected attribute. The procedure uses tree ensembles for quantile regression. Instructions for using the methods are further elaborated in the corresponding JSS manuscript, see .

fairadapt

The R package fairadapt is intended for removing bias from machine learning algorithms. In particular, it implements the pre-processing procedure described in Plecko & Meinshausen, 2019 (all the code used for producing the figures in the paper can be found in the jmlr-paper folder). The main idea is to adapt the training and testing data in a way which prevents any further training procedure from learning an undesired bias. The package currently offers the pre-processing step, after which the user can use the adapted data to train any classifier. However, some caution on the training step is still advised, so for more involved applications with resolving variables, the user should refer to the original paper.

Installation

You can install the released version of fairadapt from CRAN with:

install.packages("fairadapt")

Example

An example of how fairadapt can be used is demonstrated below on the UCI Adult dataset.

# loading the package
library(fairadapt)

vars <- c("sex", "age", "native_country", "marital_status", "education_num",
          "workclass", "hours_per_week", "occupation", "income")

# initialising the adjacency matrix
adj.mat <- c(
  0, 0, 0, 1, 1, 1, 1, 1, 1, # sex
  0, 0, 0, 1, 1, 1, 1, 1, 1, # age
  0, 0, 0, 1, 1, 1, 1, 1, 1, # native_country
  0, 0, 0, 0, 1, 1, 1, 1, 1, # marital_status
  0, 0, 0, 0, 0, 1, 1, 1, 1, # education_num
  0, 0, 0, 0, 0, 0, 0, 0, 1, # workclass
  0, 0, 0, 0, 0, 0, 0, 0, 1, # hours_per_week
  0, 0, 0, 0, 0, 0, 0, 0, 1, # occupation
  0, 0, 0, 0, 0, 0, 0, 0, 0  # income
)

adj.mat <- matrix(adj.mat, nrow = length(vars), ncol = length(vars),
                  dimnames = list(vars, vars), byrow = TRUE)

# reading in the UCI Adult data
adult <- readRDS(
  system.file("extdata", "uci_adult.rds", package = "fairadapt")
)
n <- nrow(adult) / 2

mod <- fairadapt(income ~ ., 
                 train.data = head(adult[, vars], n = n),
                 test.data = tail(adult[, vars], n = n), 
                 prot.attr = "sex", adj.mat = adj.mat, 
                 res.vars = "hours_per_week")

adapt.train <- adaptedData(mod)
adapt.test  <- adaptedData(mod, train = FALSE)

summary(mod)
#> 
#> Call:
#> fairadapt(formula = income ~ ., prot.attr = "sex", adj.mat = adj.mat, 
#>     train.data = head(adult[, vars], n = n), test.data = tail(adult[, 
#>         vars], n = n), res.vars = "hours_per_week")
#> 
#> Protected attribute:                 sex
#> Protected attribute levels:          Female, Male
#> Adapted variables:                   marital_status, education_num, workclass, occupation, income
#> Resolving variables:                 hours_per_week, age, native_country
#> 
#> Number of training samples:          1000
#> Number of test samples:              1000
#> Quantile method:                     rangerQuants
#> 
#> Total variation (before adaptation): -0.2014
#> Total variation (after adaptation):  -0.01676