Fairness.jl is a comprehensive bias audit and mitigation toolkit in julia. Extensive support and functionality provided by MLJ has been used in this package.
For an introduction to Fairness.jl refer the notebook available at https://nextjournal.com/ashryaagr/fairness
⚠️ This is experimental software: We refer Aequitas for stable bias auditing.
using Pkg
Pkg.activate("my_environment", shared=true)
Pkg.add("Fairness")
Pkg.add("MLJ")- As of writing, it is the only bias audit and mitigation toolkit to support data with multi-valued protected attribute. For eg. If the protected attribute, say race has more than 2 values: "Asian", "African", "American"..so on, then Fairness.jl can easily handle it with normal workflow.
- Multiple Fairness algorithms can be applied at the same time by wrapping the wrapped Model. Example is available in Documentation
- Due to the support for multi-valued protected attribute, intersectional fairness can also be dealt with this toolkit. For eg. If the data has 2 protected attributes, say race and gender, then Fairness.jl can be used to handle it by combining the attributes like "female_american", "male_asian"...so on.
- Extensive support and functionality provided by MLJ can be leveraged when using Fairness.jl
- Tuning of models using MLJTuning from MLJ. Numerious ML models from MLJModels can be used together with Fairness.jl
- It leverages the flexibility and speed of Julia to make it more efficient and easy-to-use at the same time
- Well structured and intutive design
- Extensive tests and Documentation
- Documentation is a good starting point for this package.
- To understand Fairness.jl, it is recommended that the user goes through the MLJ Documentation. It shall help the user in understanding the usage of machine, evaluate, etc.
- Incase of any difficulty or confusion feel free to open an issue.
Following is an introductory example of using Fairness.jl. Observe how easy it has become to measure and mitigate bias in Machine Learning algorithms.
using Fairness, MLJ
X, y, ŷ = @load_toydata
julia> model = ConstantClassifier()
ConstantClassifier() @904
julia> wrappedModel = ReweighingSamplingWrapper(classifier=model, grp=:Sex)
ReweighingSamplingWrapper(
grp = :Sex,
classifier = ConstantClassifier(),
factor = 1) @312
julia> evaluate(
wrappedModel,
X, y,
measures=MetricWrappers(
[true_positive, true_positive_rate], grp=:Sex))
┌────────────────────┬─────────────────────────────────────────────────────────────────────────────────────┬───────────────────────────────────── ⋯
│ _.measure │ _.measurement │ _.per_fold ⋯
├────────────────────┼─────────────────────────────────────────────────────────────────────────────────────┼───────────────────────────────────── ⋯
│ true_positive │ Dict{Any,Any}("M" => 2,"overall" => 4,"F" => 2) │ Dict{Any,Any}[Dict("M" => 0,"overall ⋯
│ true_positive_rate │ Dict{Any,Any}("M" => 0.8333333333333334,"overall" => 0.8333333333333334,"F" => 1.0) │ Dict{Any,Any}[Dict("M" => 4.99999999 ⋯
└────────────────────┴─────────────────────────────────────────────────────────────────────────────────────┴───────────────────────────────────── ⋯Fairness.jl is divided into following components
It is a 3D matrix of values of TruePositives, False Negatives, etc for each group. It greatly helps in optimization and removing the redundant calculations.
| Name | Metric Instances |
|---|---|
| True Positive | truepositive, true_positive |
| True Negative | truenegative, true_negative |
| False Positive | falsepositive, false_positive |
| False Negative | falsenegative, false_negative |
| True Positive Rate | truepositive_rate, true_positive_rate, tpr, recall, sensitivity, hit_rate |
| True Negative Rate | truenegative_rate, true_negative_rate, tnr, specificity, selectivity |
| False Positive Rate | falsepositive_rate, false_positive_rate, fpr, fallout |
| False Negative Rate | falsenegative_rate, false_negative_rate, fnr, miss_rate |
| False Discovery Rate | falsediscovery_rate, false_discovery_rate, fdr |
| Precision | positivepredictive_value, positive_predictive_value, ppv |
| Negative Predictive Value | negativepredictive_value, negative_predictive_value, npv |
| Name | Formula | Value for Custom function (func) |
|---|---|---|
| disparity | metric(Gᵢ)/metric(RefGrp) ∀ i | func(metric(Gᵢ), metric(RefGrp)) ∀ i |
| parity | [ (1-ϵ) <= dispariy_value[i] <= 1/(1-ϵ) ∀ i ] | [ func(disparity_value[i]) ∀ i ] |
These metrics shall use either parity or shall have custom implementation to return boolean values
| Metric | Aliases |
|---|---|
| Demographic Parity | DemographicParity |
These algorithms are wrappers. These help in mitigating bias and improve fairness.
| Algorithm Name | Metric Optimised | Supports Multi-valued protected attribute | Type | Reference |
|---|---|---|---|---|
| Reweighing | General | ✔️ | Preprocessing | Kamiran and Calders, 2012 |
| Reweighing-Sampling | General | ✔️ | Preprocessing | Kamiran and Calders, 2012 |
| Equalized Odds Algorithm | Equalized Odds | ✔️ | Postprocessing | Hardt et al., 2016 |
| Calibrated Equalized Odds Algorithm | Calibrated Equalized Odds | ❌ | Postprocessing | Pleiss et al., 2017 |
| LinProg Algorithm | Any metric | ✔️ | Postprocessing | Our own Algorithm |
| Penalty Algorithm | Any metric | ✔️ | Inprocessing | Our own Algorithm |
- Various Contribution opportunities are available. Some of the possible contributions have been listed at the pinned issue
- Feel free to open an issue or contact on slack. Let us know where your intersts and strengths lie and we can find possible contribution opportunities for you.
@software{ashrya_agrawal_2020_3977197,
author = {Ashrya Agrawal and
Jiahao Chen and
Sebastian Vollmer and
Anthony Blaom},
title = {ashryaagr/Fairness.jl},
month = aug,
year = 2020,
publisher = {Zenodo},
version = {v0.1.2},
doi = {10.5281/zenodo.3977197},
url = {https://doi.org/10.5281/zenodo.3977197}
}