Fairlearn Files

Summary

This release updates the supported Python range (drops 3.9, adds 3.13) and adds compatibility with scipy 1.16. It continues the Narwhals integration work in the postprocessing module and GroupFeature, fixes a polars dataframe-interchange deprecation, and corrects the handling of degenerate sensitive feature values in PrototypeRepresentationLearner.

Bug fixes

• Fixed handling of single-group (degenerate) sensitive features in PrototypeRepresentationLearner.fit, which previously failed during the fairness-error computation: [#1626] by @BALOGUN-DAVID.
• Changed GridSearch.fit to return self, matching the scikit-learn estimator convention: [#1622] by @k-nithin.
• Fixed a polars dataframe-interchange protocol deprecation in AnnotatedMetricFunction and corrected the 1D-array branch at the same time: [#1639] by @taharallouche.
• Added a MANIFEST.in so that requirements*.txt are included in the source distribution, and switched the build pipeline to the PEP 517 build frontend: [#1628] by @tiran.
• Excluded the test and test_othermlpackages directories from the fairlearn packaging distribution: [#1644] by @parulgupta1004.

Narwhals integration

• Made postprocessing/_tradeoff_curve_utilities.py dataframe agnostic through Narwhals: [#1553] by @FBruzzesi.
• Made GroupFeature dataframe agnostic through Narwhals: [#1546] by @FBruzzesi.

Documentation

• Added a GridSearch algorithm overview and usage example to the reductions user guide: [#1622] by @k-nithin.
• Added a usage example for mean_prediction in the advanced MetricFrame guide: [#1637] by @3Paradox.
• Improved docstrings for base metrics, DisaggregatedResult methods, and MetricFrame group methods: [#1621] by @k-nithin.
• Added versionadded and versionchanged directives to the public docstrings in fairlearn.metrics: [#1640] by @BAder82t.
• Corrected the true-positive-rate formula in the equal_opportunity_difference and equal_opportunity_ratio docstrings (from E[h(X) | A=a] to P[h(X)=1 | A=a, Y=1]): [#1605] by @OphyBoamah.
• Removed the outdated Fairlearn repository overview page from the contributor guide: [#1606] by @Anish8800.
• Fixed minor spacing and punctuation in docs/user_guide/fairness_in_machine_learning.rst: [#1611] by @Anish8800.
• Updated the Maintainers link in the README to point directly to the Maintainers section of the website instead of the general About page: [#1601] by @itz8bitstudios.

Other improvements

• Removed support for Python 3.9 and added support for Python 3.13: [#1592] by @tamaraatanasoska.
• Added support for scipy 1.16: [#1596] by @tamaraatanasoska.
• Raised the minimum supported scikit-learn to 1.5.0 and the pinned torch in requirements-min.txt to 2.8.0 to clear Dependabot security advisories: [#1645] by @romanlutz.
• Removed an unused name argument from GroupFeature: [#1599] by @taharallouche.
• Avoided a sys module patching leak in the test suite: [#1600] by @taharallouche.
• Removed a Pillow deprecation warning by requiring matplotlib>=3.10.5: [#1618] by @iclectic.
• Fixed a DataFrame fragmentation PerformanceWarning in the reductions module: [#1631] by @BALOGUN-DAVID.
• Closed Matplotlib figures before/after tests to avoid a deprecation warning: [#1603] by @lolashaffy.
• Sped up the reductions test suite: [#1574] by @taharallouche.
• Sped up the preprocessing test suite: [#1573] by @taharallouche.
• Sped up the metrics test suite: [#1572] by @taharallouche.
• Improved unit tests for PrototypeRepresentationLearner: stopped overriding parametrized sensitive_features values inside the test bodies and added coverage for multi-dimensional sensitive features: [#1623] by @taharallouche.
• Cached OpenML datasets in CI runs: [#1598] by @FBruzzesi.
• Added a Narwhals pytest marker: [#1597] by @FBruzzesi.