Fairlearn

Overview

Fairlearn is an open-source Python toolkit for assessing and improving the fairness of machine-learning models. In the wizard post it is the tool that makes algorithmic accountability “approachable” — the mechanism for auditing “the models that allocate capital.” Its homepage is fairlearn.org.

Key Facts

• Core API used in the post: MetricFrame (computes metrics disaggregated by a sensitive feature), plus the metric helpers selection_rate and demographic_parity_difference.
• Usage pattern: pass y_true, y_pred, and sensitive_features (e.g. a zip-code-derived demographic) to MetricFrame; inspect metrics.by_group; compute the demographic-parity gap as a single number.
• Sibling tool: the post names IBM’s AI Fairness 360 (AIF360) (aif360.res.ibm.com) as an equivalent alternative; both fill the same role and are not given separate treatment here.
• Reference: the post links the Fairlearn user guide under Further Reading.

Connections

• Algorithmic Accountability — the concept Fairlearn operationalizes; see that page for the full code example and decision rule.

Sources

• Wizard Topples Capitalist Dominance Ingeniously — “Example 2: Algorithmic Accountability Check.”