fairness_check_regression: Fairness check regression

In ModelOriented/FairModels: Flexible Tool for Bias Detection, Visualization, and Mitigation

Fairness check regression

Description

This is an experimental approach. Please have it in mind when using it. Fairness_check_regression enables to check fairness in regression models. It uses so-called probabilistic classification to approximate fairness measures. The metrics in use are independence, separation, and sufficiency. The intuition behind this method is that the closer to 1 the metrics are the better. When all metrics are close to 1 then it means that from the perspective of a predictive model there are no meaningful differences between subgroups.

Usage

fairness_check_regression(
  x,
  ...,
  protected = NULL,
  privileged = NULL,
  label = NULL,
  epsilon = NULL,
  verbose = TRUE,
  colorize = TRUE
)

Arguments

x	object created with explain or of class fairness_regression_object. It can be multiple fairness_objects, multiple explainers, or combination on both, as long as they predict the same data. If at least one fairness_object is provided there is no need to pass protected and privileged parameters. Explainers must be of type regression
...	possibly more objects created with explain and/or objects of class fairness_regression_object
protected	factor, protected variable (also called sensitive attribute), containing privileged and unprivileged groups
privileged	factor/character, one value of protected, denoting subgroup suspected of the most privilege
label	character, vector of labels to be assigned for explainers, default is explainer label.
epsilon	numeric, boundary for fairness checking, lowest/maximal acceptable metric values for unprivileged. Default value is 0.8.
verbose	logical, whether to print information about creation of fairness object
colorize	logical, whether to print information in color

Details

Sometimes during metric calculation faze approximation algorithms (logistic regression models) might not coverage properly. This might indicate that the membership to subgroups has strong predictive power.

References

Steinberg, Daniel & Reid, Alistair & O'Callaghan, Simon. (2020). Fairness Measures for Regression via Probabilistic Classification. - https://arxiv.org/pdf/2001.06089.pdf

Examples

set.seed(123)
data <- data.frame(
  x = c(rnorm(500, 500, 100), rnorm(500, 400, 200)),
  pop = c(rep("A", 500), rep("B", 500))
)

data$y <- rnorm(length(data$x), 1.5 * data$x, 100)

# create model
model <- lm(y ~ ., data = data)

# create explainer
exp <- DALEX::explain(model, data = data, y = data$y)

# create fobject
fobject <- fairness_check_regression(exp, protected = data$pop, privileged = "A")

# results

fobject
plot(fobject)


model_ranger <- ranger::ranger(y ~ ., data = data, seed = 123)
exp2 <- DALEX::explain(model_ranger, data = data, y = data$y)

fobject <- fairness_check_regression(exp2, fobject)

# results
fobject

plot(fobject)

ModelOriented/FairModels documentation built on Aug. 30, 2022, 5:48 p.m.