Rebalance a biased tabular dataset by group

A port of the CTGAN-style debiaser from arXiv:2603.18987. Conditions a tabular GAN on (group, outcome) and synthesises rows in which every group's favourable-outcome rate matches the privileged group's, so the Disparate-Impact Ratio of the debiased data moves toward 1.

Usage

morie_fairness_ctgan_debiaser(
  df,
  outcome_col,
  feature_cols,
  group_col = "group",
  favorable = 1L,
  privileged = NULL,
  n = 1000L,
  steps = 1500L,
  seed = 0L
)

Arguments

df

Data.frame with at least the group, outcome and feature columns.

outcome_col

Binary outcome column (favorable=1 default).

feature_cols

Character vector of continuous feature columns.

group_col

Protected-attribute column (default "group").

favorable

The favourable outcome value (default 1).

privileged

The group whose favourable rate is targeted.

n

Number of synthetic rows to return.

steps

Training iterations in the Python CTGAN; the native debiaser resamples rows rather than training a generator, so this is carried for interface parity and has no effect.

seed

Sampling/training seed.

Value

morie_fairness_result; $debiased carries the synthesised data.frame (native base-R resampling: each synthetic row draws its group from the observed group shares, its outcome at the privileged group's favourable rate, and its features from an observed row with that group and outcome).

Examples

df <- data.frame(group = c("A", "B"), outcome = c(1, 0), stringsAsFactors = FALSE)
r <- morie_fairness_ctgan_debiaser(df, outcome_col = "outcome",
  feature_cols = c("x1"), group_col = "group", n = 5L)
r$warnings
#> [1] "missing column(s): x1"