ml_fairness_check2.py

import numpy as np 

def disparate_impact(B, m, fairness_threshold=.8):
    if len(B) != len(m):
        raise ValueError('Input arrays do not have same number of entries')

    # "positive class" are those where predictions = Charge Off
    # "majority class" are those where protected class status = 1
    
    indices_pos_class, = np.where(B == 1)
    
    outcomes_pos = m[indices_pos_class]

    if len(np.where(outcomes_pos == 1)) == 0:
        return None, None
    
    value_discrim = len(np.where(outcomes_pos == 0)) / len(
        np.where(outcomes_pos == 1))

    if value_discrim <= fairness_threshold:
        is_discriminatory = True
        print("The model is DISCRIMINATING on Black with level {}".format(round(1-value_discrim,5)))
    else:
        print("The model is NOT DISCRIMINATING on Black with level {}".format(round(1-value_discrim,5)))