[FEATURE] fairness scores should work on non-binary indicators

[MBrouns]

ValueError: equal_opportunity_score only supports binary indicator columns for column. Found values ['Black' 'White']

[tbezemer]

I'm currently having a go at this. I have an implementation that passes the tests of equal_opportunity_score. I'm just not sure if my implementation of the algorithm is a valid extrapolation to the non-binary case (since I kind of based it on analysing the code alone - I need to dive into the paper).

Thoughts @MBrouns ?

Mainly the line score = ... is the part I'm unsure about.

def equal_opportunity_score(sensitive_column, positive_target=1):
    r"""
    The equality opportunity score calculates the ratio between the probability of a **true positive** outcome
    given the sensitive attribute (column) being true and the same probability given the
    sensitive attribute being false.

    .. math::
        \min \left(\frac{P(\hat{y}=1 | z=1, y=1)}{P(\hat{y}=1 | z=0, y=1)},
        \frac{P(\hat{y}=1 | z=0, y=1)}{P(\hat{y}=1 | z=1, y=1)}\right)

    This is especially useful to use in situations where "fairness" is a theme.

    Usage:
    `equal_opportunity_score('gender')(clf, X, y)`

    Source:
    - M. Hardt, E. Price and N. Srebro (2016), Equality of Opportunity in Supervised Learning

    :param sensitive_column:
        Name of the column containing the binary sensitive attribute (when X is a dataframe)
        or the index of the column (when X is a numpy array).
    :param positive_target: The name of the class which is associated with a positive outcome
    :return: a function (clf, X, y_true) -> float that calculates the equal opportunity score for z = column
    """

    def impl(estimator, X, y_true):
        """Remember: X is the thing going *in* to your pipeline."""
        sensitive_col = (
            X[:, sensitive_column] if isinstance(X, np.ndarray) else X[sensitive_column]
        )

        y_hat = estimator.predict(X)

        p_ys_zs = []

        for subgroup in np.unique(sensitive_col):
            y_given_zi_yi = y_hat[(sensitive_col == subgroup) & (y_true == positive_target)]

            # If we never predict a positive target for one of the subgroups, the model is by definition not
            # fair so we return 0
            if len(y_given_zi_yi) == 0:
                warnings.warn(
                    f"No samples with y_hat == {positive_target} for {sensitive_column} == 1, returning 0",
                    RuntimeWarning,
                )
                return 0

            p_ys_zs.append(np.mean(y_given_zi_yi == positive_target))

        #Getting the min of all pair-wise divisions is the same as getting the min of each mirror pair?
        score = np.minimum(*[ pair[0] / pair[1] for pair in it.permutations(p_ys_zs, 2)])

        return score if not np.isnan(score) else 1.

    return impl

[MBrouns]

I think it makes sense in the way you've currently written it, although it would be nice to see a few test cases to show the impact and behaviour

[tbezemer]

Thanks, good to know the logic works! I'll start working on some tests for it.