Fairness

Zhao & Welsch (2026) warn that selected rules may encode bias or act as proxies for protected attributes. flaggam.fairness operationalizes that warning: group-level performance metrics for a protected attribute, plus a rule-level audit that ranks every fitted basis by its association with it. Because FlagGAM's model is its rules, the audit inspects the actual decision logic — not a surrogate.

This module is an original addition — not part of Zhao & Welsch (2026); the design rationale is recorded in DECISIONS.

Group metrics

from flaggam import group_metrics

A = X["purpose"].astype(str)   # protected attribute (illustrative)
metrics = group_metrics(y, clf.predict_proba(X)[:, 1], A)

group_metrics reports, per level of A: n, base_rate, mean_predicted, selection_rate, tpr, auroc, and ece, plus three gap summaries (demographic_parity_diff, equal_opportunity_diff, auroc_gap) computed as the max-minus-min across groups.

Proxy audit

from flaggam import ProxyAudit

report = ProxyAudit(clf).report(X, A)                        # ranked candidate proxies
clean_clf, trade = ProxyAudit(clf).drop_proxies(X, y, A, threshold=0.3)

ProxyAudit.report ranks every fitted basis by its association with A — absolute point-biserial correlation for numeric A, Cramer's V otherwise — and flags those above threshold. ProxyAudit.drop_proxies refits only the head after removing flagged bases, returning the new estimator alongside a one-row trade-off summary (n_dropped, AUROC and demographic-parity-gap before/after), so the cost of removing a proxy is a number, not a guess.

Both methods require a fitted binary classifier with representation="full" and head="additive" (no monotonic constraints).

plot_group_metrics and plot_proxy_association visualize both outputs — see Visualization; the German Credit walkthrough runs the full audit on an illustrative sex attribute. See the API reference for the full API.