Fairlearn

Use this repo skill when a task names fairlearn, asks for the Fairlearn Python package, or uses Fairlearn vocabulary such as MetricFrame, sensitive_features, demographic parity, equalized odds, CorrelationRemover, ExponentiatedGradient, GridSearch, ThresholdOptimizer, AdversarialFairnessClassifier, or Fairlearn dataset fetchers.

Fairlearn is API-first. This checkout does not define a Fairlearn-specific command-line interface, so route future agents to Python APIs, bundled references, and the smoke scripts in this skill.

Install and quick verification

For normal package use:

python -m pip install fairlearn

Add optional plotting and adversarial-learning dependencies only when the workflow needs them:

python -m pip install matplotlib        # group plots and threshold-optimizer plots
python -m pip install torch             # PyTorch adversarial backend
# TensorFlow is an alternative adversarial backend, but was not verified in this skill run.

From this skill directory, run the bundled root check when you need a safe import and optional-backend smoke test:

python scripts/check_install.py --include-optional

Read references/installation.md for version and optional dependency details, references/troubleshooting.md for cross-cutting failures, and references/repo-provenance.md before deciding whether this skill is stale for a newer Fairlearn checkout.

Route map

Task signal	Read this sub-skill	Why
Group metrics, MetricFrame, selection_rate, demographic parity/equalized odds metrics, intersections, bootstrapped confidence intervals, model comparison plots, ROC-by-group plots	sub-skills/assessment/SKILL.md	Assessment owns Fairlearn metrics, grouped outputs, CIs, and plotting helpers.
Remove linear correlation with sensitive columns, fair representation learning, CorrelationRemover, PrototypeRepresentationLearner, preprocessing before a downstream estimator	sub-skills/preprocessing/SKILL.md	Preprocessing owns transformations applied before model training.
Fairness-constrained training with sklearn-compatible estimators, ExponentiatedGradient, GridSearch, DemographicParity, EqualizedOdds, BoundedGroupLoss, sample_weight_name	sub-skills/reductions/SKILL.md	Reductions owns algorithms that reduce fairness constraints to weighted supervised-learning problems.
Adjust a trained predictor with group-specific thresholds, ThresholdOptimizer, plot_threshold_optimizer, prefit, predict_method, post-fit parity constraints	sub-skills/postprocessing/SKILL.md	Postprocessing owns prediction-time threshold/interpolation mitigation.
Neural adversarial fairness, PyTorch or TensorFlow backends, callbacks, CUDA, AdversarialFairnessClassifier, AdversarialFairnessRegressor	sub-skills/adversarial/SKILL.md	Adversarial workflows have backend-specific model, optimizer, and training pitfalls.
Built-in datasets, fetch_adult, fetch_acs_income, fetch_boston, return_X_y, as_frame, data_home, dataset fairness caveats	sub-skills/datasets/SKILL.md	Dataset loaders have network/cache behavior and dataset-specific schemas.
Installing, checking versions, optional extras, missing matplotlib/torch/tensorflow, show_versions, import/runtime diagnostics	sub-skills/installation/SKILL.md	Installation owns environment and dependency recovery paths.

Operating rules

• Treat Fairlearn as a sociotechnical fairness toolkit, not an automated fairness oracle. Metric choices and mitigation goals must be tied to the user's problem context.
• Keep row alignment explicit. X, y_true/y, predictions, and sensitive_features must describe the same samples in the same order.
• Use sensitive_features for group definitions even when those columns are also present in X. Do not assume ignoring sensitive features is enough to assess or mitigate unfairness.
• Start with assessment before mitigation unless the user already gives a constrained mitigation algorithm and metric target.
• Prefer tiny synthetic smoke checks or cached data before networked dataset downloads. Fairlearn dataset fetchers use OpenML-style downloads and cache under a data home.
• Make optional dependencies explicit. Plotting requires matplotlib; adversarial mitigation requires either PyTorch or TensorFlow. The PyTorch CPU path and optional CUDA acceleration were verified for this skill; TensorFlow was only documented as an alternative.
• The generated references and scripts are self-contained. Do not instruct future agents to open this repo's original docs, tests, examples, or source scripts for normal operation.

Shared runtime files

• references/fairness-framing.md summarizes Fairlearn's group-fairness framing, sensitive features, parity constraints, disparity metrics, and limits.
• references/installation.md covers install commands, optional dependencies, import checks, and version inspection.
• references/troubleshooting.md covers package-wide import, plotting, backend, data-shape, and stale-version failures.
• references/repo-provenance.md records source revision, package version, evidence paths, backend verification, and refresh signals.
• references/repo-routing-metadata.json is structured metadata consumed by DisCo's managed repo-skills router import transaction.
• scripts/check_install.py validates the installed Fairlearn package, stable public submodules, and optional plotting/adversarial dependencies.

Good first branch decisions

• If the prompt says "is my model fair?", "compare groups", or names a metric, start with assessment.
• If it asks to transform features before training, start with preprocessing.
• If it asks to train a fair model from scratch with a sklearn estimator, start with reductions.
• If it already has a fitted predictor or score function and wants fair thresholds, start with postprocessing.
• If it names neural networks, PyTorch, TensorFlow, callbacks, or CUDA, start with adversarial.
• If it asks for Adult, ACSIncome, Boston, diabetes, bank, or credit-card sample data, start with datasets.
• If imports, missing extras, or version questions block the task, start with installation and root troubleshooting.