Skip to main content

Algorithms for mitigating unfairness in supervised machine learning

Project description

Build Status MIT license PyPI Gitter StackOverflow


Fairlearn is a Python package that empowers developers of artificial intelligence (AI) systems to assess their system's fairness and mitigate any observed unfairness issues. Fairlearn contains mitigation algorithms as well as a Jupyter widget for model assessment. Besides the source code, this repository also contains Jupyter notebooks with examples of Fairlearn usage.

Current release

  • The current stable release is available at Fairlearn v0.4.6.

  • Our current version differs substantially from version 0.2 or earlier. Users of these older versions should visit our onboarding guide.

What we mean by fairness

An AI system can behave unfairly for a variety of reasons. In Fairlearn, we define whether an AI system is behaving unfairly in terms of its impact on people – i.e., in terms of harms. We focus on two kinds of harms:

  • Allocation harms. These harms can occur when AI systems extend or withhold opportunities, resources, or information. Some of the key applications are in hiring, school admissions, and lending.

  • Quality-of-service harms. Quality of service refers to whether a system works as well for one person as it does for another, even if no opportunities, resources, or information are extended or withheld.

We follow the approach known as group fairness, which asks: Which groups of individuals are at risk for experiencing harms? The relevant groups need to be specified by the data scientist and are application specific.

Group fairness is formalized by a set of constraints, which require that some aspect (or aspects) of the AI system's behavior be comparable across the groups. The Fairlearn package enables assessment and mitigation of unfairness under several common definitions. To learn more about our definitions of fairness, please visit our user guide on Fairness of AI Systems.

Note: Fairness is fundamentally a sociotechnical challenge. Many aspects of fairness, such as justice and due process, are not captured by quantitative fairness metrics. Furthermore, there are many quantitative fairness metrics which cannot all be satisfied simultaneously. Our goal is to enable humans to assess different mitigation strategies and then make trade-offs appropriate to their scenario.

Overview of Fairlearn

The Fairlearn package has two components:

  • A dashboard for assessing which groups are negatively impacted by a model, and for comparing multiple models in terms of various fairness and accuracy metrics.

  • Algorithms for mitigating unfairness in a variety of AI tasks and along a variety of fairness definitions.

Fairlearn algorithms

Fairlearn contains the following algorithms for mitigating unfairness in binary classification and regression:

algorithm description classification/regression sensitive features supported fairness definitions
fairlearn. reductions. ExponentiatedGradient Black-box approach to fair classification described in A Reductions Approach to Fair Classification binary classification categorical DP, EO
fairlearn. reductions. GridSearch Black-box approach described in Section 3.4 of A Reductions Approach to Fair Classification binary classification binary DP, EO
fairlearn. reductions. GridSearch Black-box approach that implements a grid-search variant of the algorithm described in Section 5 of Fair Regression: Quantitative Definitions and Reduction-based Algorithms regression binary BGL
fairlearn. postprocessing. ThresholdOptimizer Postprocessing algorithm based on the paper Equality of Opportunity in Supervised Learning. This technique takes as input an existing classifier and the sensitive feature, and derives a monotone transformation of the classifier's prediction to enforce the specified parity constraints. binary classification categorical DP, EO

Note: DP refers to demographic parity, EO to equalized odds, and BGL to bounded group loss. For more information on these and other terms we use in this repository please refer to the terminology page. To request additional algorithms or fairness definitions, please open a new issue.

Fairlearn dashboard

Fairlearn dashboard is a Jupyter notebook widget for assessing how a model's predictions impact different groups (e.g., different ethnicities), and also for comparing multiple models along different fairness and accuracy metrics.

Set-up and a single-model assessment

To assess a single model's fairness and accuracy, the dashboard widget can be launched within a Jupyter notebook as follows:

from fairlearn.widget import FairlearnDashboard

# A_test containts your sensitive features (e.g., age, binary gender)
# sensitive_feature_names containts your sensitive feature names
# y_true contains ground truth labels
# y_pred contains prediction labels

                   sensitive_feature_names=['BinaryGender', 'Age'],

After the launch, the widget walks the user through the assessment set-up, where the user is asked to select (1) the sensitive feature of interest (e.g., binary gender or age), and (2) the accuracy metric (e.g., model precision) along which to evaluate the overall model performance as well as any disparities across groups. These selections are then used to obtain the visualization of the model's impact on the subgroups (e.g., model precision for females and model precision for males).

The following figures illustrate the set-up steps, where binary gender is selected as a sensitive feature and accuracy rate is selected as the accuracy metric.

Dashboard set-up

After the set-up, the dashboard presents the model assessment in two panels:

Panel Description
Disparity in accuracy This panel shows: (1) the accuracy of your model with respect to your selected accuracy metric (e.g., accuracy rate) overall as well as on different subgroups based on your selected sensitive feature (e.g., accuracy rate for females, accuracy rate for males); (2) the disparity (difference) in the values of the selected accuracy metric across different subgroups; (3) the distribution of errors in each subgroup (e.g., female, male). For binary classification, the errors are further split into overprediction (predicting 1 when the true label is 0), and underprediction (predicting 0 when the true label is 1).
Disparity in predictions This panel shows a bar chart that contains the selection rate in each group, meaning the fraction of data classified as 1 (in binary classification) or distribution of prediction values (in regression).

Fairness Insights

Comparing multiple models

The dashboard also enables comparison of multiple models, such as the models produced by different learning algorithms and different mitigation approaches, including fairlearn.reductions.GridSearch, fairlearn.reductions.ExponentiatedGradient and fairlearn.postprocessing.ThresholdOptimizer.

As before, the user is first asked to select the sensitive feature and the accuracy metric. The model comparison view then depicts the accuracy and disparity of all the provided models in a scatter plot. This allows the user to examine trade-offs between accuracy and fairness. Each of the dots can be clicked to open the assessment of the corresponding model. The figure below shows the model comparison view with binary gender selected as a sensitive feature and accuracy rate selected as the accuracy metric.

Accuracy Fairness Tradeoff

Install Fairlearn

The package can be installed via

pip install fairlearn

or optionally with a full feature set by adding extras, e.g. pip install fairlearn[customplots].

or you can clone the repository locally via

git clone

To verify that the cloned repository works (the pip package does not include the tests), run

pip install -r requirements.txt
python -m pytest -s ./test/unit
Onboarding guide for users of version 0.2 or earlier

Up to version 0.2, Fairlearn contained only the exponentiated gradient method. The Fairlearn repository now has a more comprehensive scope and aims to incorporate other methods as specified above. The same exponentiated gradient technique is now the class fairlearn.reductions.ExponentiatedGradient. While in the past exponentiated gradient was invoked via

import numpy as np
from fairlearn.classred import expgrad
from fairlearn.moments import DP

estimator = LogisticRegression()  # or any other estimator
exponentiated_gradient_result = expgrad(X, sensitive_features, y, estimator, constraints=DP())
positive_probabilities = exponentiated_gradient_result.best_classifier(X)
randomized_predictions = (positive_probabilities >= np.random.rand(len(positive_probabilities))) * 1

the equivalent operation is now

from fairlearn.reductions import ExponentiatedGradient, DemographicParity

estimator = LogisticRegression()  # or any other estimator
exponentiated_gradient = ExponentiatedGradient(estimator, constraints=DemographicParity()), y, sensitive_features=sensitive_features)
randomized_predictions = exponentiated_gradient.predict(X)

Please open a new issue if you encounter any problems.


For common usage refer to the Jupyter notebooks and our API guide


To contribute please check our contributing guide.


The Fairlearn project is maintained by:

  • @MiroDudik
  • @riedgar-ms
  • @rihorn2
  • @romanlutz

For a full list of contributors refer to the authors page


Usage Questions

Pose questions and help answer them on Stack Overflow with the tag fairlearn or on Gitter.

Regular (non-security) issues

Issues are meant for bugs, feature requests, and documentation improvements. Please submit a report through GitHub issues. A maintainer will respond promptly as appropriate.

Maintainers will try to link duplicate issues when possible.

Reporting security issues

Please take a look at our guidelines for reporting security issues.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for fairlearn, version 0.4.6
Filename, size File type Python version Upload date Hashes
Filename, size fairlearn-0.4.6-py3-none-any.whl (21.2 MB) File type Wheel Python version py3 Upload date Hashes View
Filename, size fairlearn-0.4.6.tar.gz (10.5 MB) File type Source Python version None Upload date Hashes View

Supported by

Pingdom Pingdom Monitoring Google Google Object Storage and Download Analytics Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page