Fairness Checker

This Python module fairness_checker provides a set of methods to evaluate the fairness of a predictive model's outcomes across different demographic groups represented in a CSV file or given a model.

Dependencies

• Python >= 3.8

Installation

pip3 install fairness-checker

Usage

As a library

CSV checker

First set up the checker using a benchmark dataset:

from fairness_checker import fairness_csv_checker
c = fairness_csv_checker("compas-scores-two-years.csv")

Then you can call fairness measure functions. For example:

c.demographic_parity(0.2, lambda row: row['sex'] == 'Male', lambda row: row['score_text'] in {'Medium', 'High'})

Output:

demographic parity
fair: 0.04 < 0.2

Note the function signature of demographic_parity:

demographic_parity(ratio: float,
                   privileged_predicate: Callable[[csv_row], bool],
                   positive_predicate: Callable[[T], bool]) -> bool:

Here the privileged_predicate is

lambda row: row['sex'] == 'Male'

meaning the privileged group is the male group, and the positive_predicate is

lambda row: row['score_text'] in {'Medium', 'High'}

meaning the row is positive if the score is categorized as medium or high.

For a more complicated example involving parameters:

c.conditional_statistical_parity(0.2, lambda row: row['sex'] == sex, lambda row: row['score_text'] in {'Medium', 'High'}, lambda x: (lambda row: int(row['priors_count']) > x), (0,))

Output:

conditional statistical parity
fair: 0.04 < 0.2

Note the function signature of conditional_statistical_parity:

def conditional_statistical_parity(ratio: float,
                                   privileged_predicate: Callable[[csv_row], bool],
                                   positive_predicate: Callable[[csv_row], bool],
                                   legitimate_predicate_h: Callable[..., Callable[[csv_row], bool]],
                                   legitimate_arg: Tuple[Any, ...]) -> bool:

Here the higher order function legitimate_predicate_h is

lambda x: (lambda row: int(row['priors_count']) > x)

and the argument to it, legitimate_arg, is (0,).

Model checker

from fairness_checker import fairness_model_checker
c = fairness_model_checker("compas-scores-two-years.csv")

Alternatively, you can use the checker on a model. It expects the model to have a predict method that takes a csv filename as input and returns an iterable of results.

c.demographic_parity(0.2, model, lambda row: row['sex'] == 'Male', lambda Y: Y == 1)

The last predicate here is used on the model result.

As a command line CLI

Prepare your dataset file. Create a predicate definition file containing arguments to the measure functions. For example, to calculate negative balance, create a file test_predicates1.py containing the following:

def privileged_predicate(row):
    return row['sex'] == 'Male'

def score_predicate(row):
    return int(row['decile_score'])

def truth_predicate(row):
    return row['is_recid'] == '1'

Make sure the order of the definitions are the same as the order of the function signature.

Then execute the client in command line:

python3 -m fairness_checker

You'll be asked a few questions about which fairness measure you want to calculate and what ratio you want to set, like so:

Input dataset file name: compas-scores-two-years.csv
Input ratio: 0.2
Input the fairness measure: negative balance
Input the predicate definitions file name: test_predicates1.py

Output:

negative balance
fair: 0.07 < 0.2

For another example, let's calculate equal calibration with a predicate file test_predicates2.py containing the following:

def privileged_predicate(row):
    return row['sex'] == 'Male'

def truth_predicate(row):
    return row['is_recid'] == '1'

def calib_predicate_h(u, l):
    def tmp(row):
        return l <= int(row['decile_score']) and int(row['decile_score']) <= u
    return tmp

calib_arg = (7, 5)

Again, the order of the definition matters. They must match that of the function signature.

Execute in command line:

python3 -m fairness_checker

You'll be asked a few questions about which fairness measure you want to calculate and what ratio you want to set, like so:

Input dataset file name: compas-scores-two-years.csv
Input ratio: 0.2
Input the fairness measure: equal calibration
Input the predicate definitions file name: test_predicates2.py

Output:

equal calibration
fair: 0.10 < 0.2