Fair Gated Algorithm for Targeted Equity

Fairness-aware correction of a prediction model for binary and continuous outcomes with two or more protected groups and their intersections. A cross-fitted base learner (gradient-boosted trees or a neural network) is fitted without the protected attribute, and a gated mixture of neural experts learns a per-person correction to its score under a fairness penalty (a random-effects or variance penalty on outcome-stratified group means, or a thresholded worst-group penalty on error rates), with a cross-fitted disparity signal, an intervention cost, smoothed early stopping, a per-person export of adjustments with bootstrap stability intervals, and routing analyses of which expert corrects whom. Methods draw on mixture-of-experts neural networks by Jordan and Jacobs (1994) , gradient boosting by Chen and Guestrin (2016) , cross-fitting by Chernozhukov and co-authors (2018) , equalised odds by Hardt, Price and Srebro (2016) , the random-effects variance estimator of DerSimonian and Laird (1986) , and personalised treatment prediction for depression by Iniesta, Stahl and McGuffin (2016) .

fairGATE

fairGATE (Fair Gated Algorithm for Targeted Equity) corrects an existing prediction model so that its errors are spread more evenly across protected groups, and reports, person by person, how far and how reliably each prediction was moved. It works for binary and continuous outcomes, for two or more groups and for intersections of two attributes, in any domain where a prediction is used to make decisions about people: health care, income and credit, or the review of expert judgements.

Developed by Rhys Holland under the supervision of Dr Raquel Iniesta, King's College London (https://www.kcl.ac.uk/people/raquel-iniesta).

Installation

install.packages("fairGATE")
# development version
# remotes::install_github("rhysholland/FairGATE")

# the neural-network components need the LibTorch backend once:
torch::install_torch()

The base learner uses the xgboost package (base_learner = "xgboost") or torch (base_learner = "nn"); both are suggested dependencies.

How it works

A base learner is fitted without the protected attribute and cross-fitted: training rows get out-of-fold scores, held-out rows the prediction of a fit on all training rows (fairgate_base_learner(); XGBoost or a neural network, each with its own tuning grid).
A gate weights the base learner and a set of neural experts for each person; each expert proposes a correction to the base score. The untrained model is exactly the base learner.
The correction is trained under a fairness penalty: a random-effects or variance penalty on outcome-stratified group means, or a thresholded worst-group penalty on error rates, plus an intervention cost for moving people away from the base score. A cross-fitted disparity signal tells the gate where corrections would close group gaps; the protected attribute itself is never an input.
A configuration is chosen on an inner split by one rule: the smallest worst-group disparity among configurations within 0.01 of the best validation AUC or R-squared.
Every held-out person receives a base score, a corrected score, the adjustment, the routing weights and a bootstrap stability interval for the adjustment (fairgate_export()).

Quick start

library(fairGATE)
torch::torch_set_num_threads(1)          # reproducible fits

set.seed(1)
n <- 2000
g <- sample(0:2, n, replace = TRUE, prob = c(0.6, 0.25, 0.15))
X <- cbind(x1 = rnorm(n) + 0.5 * g, x2 = rnorm(n), x3 = rnorm(n))
y <- rbinom(n, 1, plogis(X[, 1] - 0.5 * X[, 2]))

fit <- fairgate(X, y, group = g, family = "binomial", reference_group = 0,
                base_learner = "xgboost")   # or "nn"
fit
head(fairgate_export(fit))

fairgate() uses the full protocol of the accompanying study by default (48 searched configurations, up to 300 epochs, 10 bootstrap replicates), which takes minutes on tens of thousands of rows. For a first look pass, for example, n_draws = 4, epochs = 50, interval_reps = 2.

The building blocks are exported for custom pipelines: fairgate_base_learner() and fairgate_tune_base_learner() for the base learner, train_gnn() for the correction with supplied base scores, fairgate_disparity_target(), fairgate_group_disparity() and the penalty functions, and analyse_experts(), analyse_gnn_results(), plot_sankey() and export_f360_csv() (AI Fairness 360 format) for analysis.

See vignette("introduction-to-fairGATE") for a worked example on the Adult income sample shipped with the package.