Downstream Fairness
We implement the algorithm presented in Geometric Repair for Fair Classification at Any Decision Threshold.
The algorithm looks at the scores, labels, and sensitive attribute in a dataset, and determines a set of adjustments.
The original set of adjustments allow one to adjust the prediction probabilities of a binary classifier, and perturb
them (by adding the appropriate adjustment to a prediction probability), so that you can achieve demographic parity.
To achieve other fairness definitions, such as equalized odds and equal opportunity, our algorithm provides a lambda
value. This lambda value is multiplied to each entry of the set of adjustments, and then those can be added to
predictions. Here is a good way to think about it:
prediction_probability + adjustment_value # achieves demographic parity
prediction_probability + lambda_val * adjustment_value # achieves some fairness definition attached to the lambda value
Quickstart
First, please make sure to download our package. You can either do this from this repo via:
pip install -e .
or you can install it directly via pip:
pip install downstream_fairness
To get started, we recommend just running our get_bias_mitigator, so you can get a dictionary of lambda values and
an adjustment table. To go more in-depth, we recommend going through our demo notebook.
from downstream_fairness.process import get_bias_mitigator
# Gets the adjustment table and lambdas
table, lambdas = get_bias_mitigator(YOUR_DATA,
sens_col=YOUR_SENSITIVE_ATTRIBUTE_COLUMN_NAME,
score_col=YOUR_PREDICTION_PROBABILTIY_COLUMN,
label_col=YOUR_LABEL_COLUMN)
# Adjusts the prediction probabilities on the fly
adjusted_scores = get_adusted_scores(table,
DATA_YOU_WANT_ADJUSTED,
YOUR_SENSITIVE_ATTRIBUTE_COLUMN_NAME,
YOUR_PREDICTION_PROBABILTIY_COLUMN,
lambdas[THE_METRIC_YOU_WANT])