DSLD supports two related tasks: estimating differences between sensitive groups and limiting sensitive-attribute influence in predictive models.
| Area | Functions |
|---|---|
| Adjusted group comparisons | dsldLinear, dsldLogit,
dsldML |
| Confounder and proxy exploration | dsldCHunting, dsldOHunting,
dsldTakeALookAround |
| Group frequency and density summaries | dsldFrequencyByS, dsldDensityByS,
dsldConfounders |
| Multivariate graphics | dsldFreqPCoord, dsldScatterPlot3D,
dsldConditDisparity |
| Matching and structure discovery | dsldMatchedATE, dsldIamb |
| fairml models | dsldFrrm, dsldFgrrm,
dsldNclm, dsldZlm, dsldZlrm |
| Explicitly deweighted features | dsldQeFairKNN, dsldQeFairRF,
dsldQeFairRidgeLin, dsldQeFairRidgeLog |
| Cross-validated parameter comparison | dsldFairUtils |
Use summary() for the linear/logistic group comparisons
and predict() for new data. Each function’s help page
describes its arguments and return values.
The fair-ML fields trainAcc and testAcc are
prediction errors, despite their historical names: mean absolute error
for regression and misclassification rate for binary classification.
Smaller errors indicate better prediction. Signed correlations describe
association with each sensitive attribute or factor level; they are not
a universal fairness guarantee.
The optional packages Matching, bnlearn,
randomForest, cccp, and CVXR are
required only by their corresponding methods. Consult each function’s
help page for its inputs, return values, and examples.