Fairlearn can be installed with pip from PyPI as follows:

pip install fairlearn

Fairlearn is also available on conda-forge:

conda install -c conda-forge fairlearn

For checking out the latest version in our repository check out our advanced_install. If you are updating from a previous version of Fairlearn, please see Migrating from prior versions.


The Fairlearn API is still evolving, so example code in this documentation may not work with every version of Fairlearn. Please use the version selector to get to the instructions for the appropriate version. The instructions for the main branch require Fairlearn to be installed from a clone of the repository. See advanced_install for the required steps.

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 in 10 minutes

The Fairlearn tookit can assist in assessing and mitigation unfairness in Machine Learning models. It’s impossible to provide a sufficient overview of fairness in ML in this Quickstart tutorial, so we highly recommend starting with our User Guide. Fairness is a fundamentally sociotechnical challenge and cannot be solved with technical tools alone. They may be helpful for certain tasks such as assessing unfairness through various metrics, or to mitigate observed unfairness when training a model. Additionally, fairness has different definitions in different contexts and it may not be possible to represent it quantitatively at all.

Given these considerations this Quickstart tutorial merely provides short code snippet examples of how to use basic Fairlearn functionality for those who are already intimately familiar with fairness in ML. The example below is about binary classification, but we similarly support regression.

Loading the dataset

For this example we use the UCI adult dataset where the objective is to predict whether a person makes more (label 1) or less (0) than $50,000 a year.

>>> import numpy as np
>>> import pandas as pd
>>> import matplotlib.pyplot as plt
>>> from sklearn.datasets import fetch_openml
>>> data = fetch_openml(data_id=1590, as_frame=True)
>>> X = pd.get_dummies(data.data)
>>> y_true = (data.target == '>50K') * 1
>>> sex = data.data['sex']
>>> sex.value_counts()
Male      32650
Female    16192
Name: sex, dtype: int64

Mitigating disparity

If we observe disparities between groups we may want to create a new model while specifying an appropriate fairness constraint. Note that the choice of fairness constraints is crucial for the resulting model, and varies based on application context. If selection rate is highly relevant for fairness in this contrived example, we can attempt to mitigate the observed disparity using the corresponding fairness constraint called Demographic Parity. In real world applications we need to be mindful of the sociotechnical context when making such decisions. The Exponentiated Gradient mitigation technique used fits the provided classifier using Demographic Parity as the objective, leading to a vastly reduced difference in selection rate:

>>> from fairlearn.reductions import ExponentiatedGradient, DemographicParity
>>> np.random.seed(0)  # set seed for consistent results with ExponentiatedGradient
>>> constraint = DemographicParity()
>>> classifier = DecisionTreeClassifier(min_samples_leaf=10, max_depth=4)
>>> mitigator = ExponentiatedGradient(classifier, constraint)
>>> mitigator.fit(X, y_true, sensitive_features=sex)
>>> y_pred_mitigated = mitigator.predict(X)
>>> sr_mitigated = MetricFrame(selection_rate, y_true, y_pred_mitigated, sensitive_features=sex)
>>> print(sr_mitigated.overall)
>>> print(sr_mitigated.by_group)
Female    0.1552...
Male      0.1715...
Name: selection_rate, dtype: object

Similarly, we can explore the difference between the initial model and the mitigated model with respect to selection rate and accuracy in the dashboard through a multi-model comparison:

>>> FairlearnDashboard(sensitive_features=sex,
...                    sensitive_feature_names=['sex'],
...                    y_true=y_true,
...                    y_pred={"initial model": y_pred, "mitigated model": y_pred_mitigated}) 

What’s next?

Please refer to our User Guide for a comprehensive view on Fairness in Machine Learning and how Fairlearn fits in, as well as an exhaustive guide on all parts of the toolkit. For concrete examples check out the Example Notebooks section. Finally, we also have a collection of Frequently asked questions.