Counterfactual fairness via GAN-style generative models
Source:R/fairness_gan.R
morie_fairness_gan.RdR ports of the JAX GAN primitives in morie.fairness.gan.
Two callables: morie_fairness_spatial_gan learns a 2-D
coordinate distribution and samples synthetic points;
morie_fairness_ctgan_debiaser rebalances a tabular dataset
so every group's favourable-outcome rate matches a privileged
group's. Both run on a native base-R backend on every install: the
spatial generator is a small leaky-ReLU MLP fitted by a seeded
random-search (keep-if-better) loop, and the debiaser resamples
observed rows conditional on (group, outcome). Neither torch nor
reticulate/JAX is used.