dsld (version 0.2.2)

dsldFairUtilTrade: dsldFairUtilTrade

Description

Exploration of the Fairness-Utility Tradeoff. Finds predictive accuracy and correlation between S and predicted Y.

Usage

dsldFairUtilTrade(data,yName,sName,dsldFtnName,
   unfairness=NULL,deweightPars=NULL,yesYVal=NULL,yesSVal=NULL,
   corrType='kendall', holdout = floor(min(1000, 0.1 * nrow(data))))

Value

A two-component vector, consisting of predictive accuracy and strength of relation between S and predicted Y.

Arguments

data

Data frame.

yName

Name of the response variable Y column. Y must be numeric or binary (two-level R factor).

sName

Name of the sensitive attribute S column. S must be numeric or binary (two-level R factor).

dsldFtnName

Quoted name of one of the fairML or EDF functions.

unfairness

Nonnull for the fairML functions.

deweightPars

Nonnull for the EDF functions.

yesYVal

Y value to be treated as Y = 1 for binary Y.

yesSVal

S value to be treated as S = 1 for binary S.

corrType

Either 'kendall' or 'probs'.

holdout

Size of holdout set.

Author

N. Matloff

Details

Tool for exploring tradeoff between utility (predictive accuracy, Mean Absolute Prediction Error or overall probability of misclassification) and fairness. Roughly speaking, the latter is defined as the strength of relation between S and predicted Y (the smaller, the better). The main issue is definition of "relation" in the case of binary Y or S:

In the 'kendall' case, binary predicted Y or S is recoded to 1s and 0s, and Kendall correlation is used. In the 'probs' case, binary Y or S is replaced by P(Y = 1 | X) and P(S = 1 | X); squared Pearson correlation is then computed.

Examples

data(svcensus)
dsldFairUtilTrade(svcensus,'wageinc','gender','dsldFrrm',0.2,yesSVal='male')
data(lsa)
race1 <- lsa$race1
lsabw <- lsa[race1 == 'black' | race1 == 'white',]
# need to get rid of excess levels
race1 <- lsabw$race1
race1 <- as.character(race1)
lsabw$race1 <- as.factor(race1)
dsldFairUtilTrade(lsabw,'bar','race1','dsldQeFairRidgeLog',
   deweightPars=list(fam_inc=0.1),yesYVal='TRUE',yesSVal='white')