Fairness

Fairness.jl is a bias audit and mitigation toolkit in julia and is supported by MLJ Ecosystem.

Installation

julia> using Pkg
julia> Pkg.activate("my_environment", shared=true)
julia> Pkg.add("Fairness")
julia> Pkg.add("MLJ")

Components

It shall be divided into following components

• FairTensor
• Measures
  • CalcMetrics
  • FairMetrics
  • BoolMetrics
• Algorithms
  • Preprocessing Algorithms
  • InProcessing Algorithms
  • PostProcessing Algorithms
• Fairness Datasets (Macros)

Important Points to Note

After you go through the documentation or have a basic idea of the package, revisit the following points which are essential to make best out this package.

• Almost every fairness dataset has a categorical field. To be able to use the various MLJ Models, you should also use a ContinuousEncoder in the following manner.

model = @pipeline ContinuousEncoder @load(SomeClassifier, pkg=PackageOfClassifier)

• To get a list of classifiers that can be used with Fairness.jl, execute

using MLJ
models(x->x.target_scitype<:AbstractVector{<:Finite})

After you get the list, pick a classifier of your choice. Lets say you choose the tuple with (name=RandomForestClassifier, package_name=DecisionTree). Then to use the model, you have to execute

model = @pipeline ContinuousEncoder @load(RandomForestClassifier, pkg=DecisionTree)

Note that you might be asked to install a specific package for the classifier. Execute the instruction of the type import Pkg; Pkg.add("--"), which you will be asked to do when you try to load the model.