SSFairnessAudit

SSFairnessAudit is an R package for auditing group fairness metrics when labels are fully observed or only partially available. The repository contains the package source, generated documentation, and a small set of research scripts used for simulation experiments.

Main workflow

Use Audit_Fairness() as the high-level entry point or call SSFairness() directly when you want more control over the semi-supervised estimator.

SSFairness() now supports useful controls:

• cross_fit_variance = TRUE to use the cross-fitted imputation path for variance estimation
• return_imputation_quality = TRUE to return imputation diagnostics plus the labeled and unlabeled imputations
• folds = ... to reuse the same labeled-data folds across candidate models when comparing them with Select_Model()

The semi-supervised basis options now include polynomial, natural spline, interaction, beta-calibration, and kernel branches. The natural spline path is available through basis = "Spline(S)", basis = "Spline(S) + X", and basis = "Spline Interaction". The additive spline branch uses a shared smooth in S plus additive covariate effects; the spline interaction branch adds spline-by-covariate interactions so the shape in S can vary with X.

Repository layout

• R/: package source code
• man/: generated .Rd documentation
• docs/: pkgdown site output
• scripts/: simulation and exploratory analysis scripts

Main user-facing functions

• Audit_Fairness(): wrapper for supervised and semi-supervised auditing
• SupervisedFairness(): fairness estimation with labeled outcomes
• SSFairness(): semi-supervised fairness estimation and optional imputation diagnostics
• DataGeneration(): synthetic data generator for simulations
• Select_Model(): candidate-model selection helper Default selection now uses a TPR-weighted cross-fitted squared-error criterion, with criterion = "brier" still available for the plain Brier score.

When comparing candidate semi-supervised models with cross-fitted imputation quality, reuse the same folds object across all SSFairness() calls so the comparison is based on the same labeled-data splits.

Notes

• docs/ is committed so the package website can be served from GitHub Pages.
• scripts/ is intentionally excluded from package builds because it supports experiments rather than the package API.