fairlearn

What it is and what it does

Fairlearn is a Python package for assessing and mitigating fairness in AI systems, focusing on harms that occur when models treat different groups of people unequally. It operationalizes fairness through a group fairness approach, asking which groups are at risk for experiencing harms—either allocation harms (when opportunities or resources are withheld) or quality-of-service harms (when a system works differently for different people). The package provides two main components: metrics for identifying which groups are negatively impacted by a model and comparing models across fairness and accuracy trade-offs, and mitigation algorithms that reduce unfairness across various AI tasks and fairness definitions.

The package depends on standard data science libraries (numpy, pandas, scikit-learn, scipy) and narwhals, and supports Python 3.10 through 3.13. It is actively maintained with a release history dating back to 2018, and the repository shows recent activity. The package acknowledges that fairness is fundamentally a sociotechnical challenge—quantitative metrics alone cannot capture all aspects like justice or due process, and not all fairness metrics can be satisfied simultaneously. Its role is to enable practitioners to assess different mitigation strategies and make trade-offs appropriate to their specific scenario.

Use it for

• Assess hiring or lending models to detect whether they systematically disadvantage protected groups before deployment
• Measure and compare fairness metrics across multiple candidate models to inform model selection decisions
• Apply post-processing mitigation algorithms to adjust model predictions to satisfy fairness constraints like demographic parity
• Evaluate quality-of-service fairness to ensure a recommendation or classification system works equally well across demographic groups
• Generate fairness reports for stakeholders to document which groups experience harms and what trade-offs were made

Worth the install?

AI-flagged interpretation of the facts on this page. Verify before relying on it.

Install

fairlearn on PyPI

Before you install

Low install friction with a pure Python wheel and five well-established dependencies (numpy, pandas, scikit-learn, scipy, narwhals). Active maintenance with a release 68 days ago and ongoing repository activity.

License in practice

Permissive license allows unrestricted use, modification, and distribution in commercial and private projects without copyleft obligations.

Quickstart

pip install fairlearn

from fairlearn.metrics import MetricFrame
from fairlearn.postprocessing import ThresholdOptimizer
import pandas as pd

# Assess fairness across groups
metric_frame = MetricFrame(metrics={...}, y_true=y, y_pred=y_pred, groups=groups)

# Mitigate unfairness
mitigator = ThresholdOptimizer(estimator=model, constraints='demographic_parity')
mitigator.fit(X_train, y_train, sensitive_features=sensitive_train)

Package facts

Evidence: fairlearn-0.14.0-py3-none-any.whl

Tags

More Artificial Intelligence packages

See also empirical-calibration · arthur-client · tslearn · pyannote-metrics · pyiqa · moocore · imagededup · river · learnosity-sdk · psmpy

Further reading