fatf.fairness.models.measures.demographic_parity

fatf.fairness.models.measures.demographic_parity(confusion_matrix_list: List[numpy.ndarray], tolerance: float = 0.2, label_index: int = 0) → numpy.ndarray

Checks for demographic parity between all of the sub-populations.

This function checks if predictive positive rate difference of all grouping pairs is within the tolerance level.

Note

This function expects a list of confusion matrices per sub-group for tested data. To get this list please use either fatf.utils.metrics.tools.confusion_matrix_per_subgroup or fatf.utils.metrics.tools.confusion_matrix_per_subgroup_indexed function.

Alternatively you can call either fatf.fairness.models.measures.disparate_impact or fatf.fairness.models.measures.disparate_impact_indexed function, which handles both the grouping and calculates the desired group-fairness criterion.

Parameters

confusion_matrix_listList[numpy.ndarray]

A list of confusion matrices, one for each sub-population.

tolerancenumber, optional (default=0.2)

A number between 0 and 1 that indicates how much any two predictive positive rates can differ to be considered “equal”.

label_indexinteger, optional (default=0)

The index of the “positive” class in the confusion matrix. (Not required for binary problems.)

Returns

disparitynumpy.ndarray

A square and diagonally symmetric numpy array with boolean values. An entry is True if a pair of two sub-populations’ predictive positive rate difference is above the tolerance level and False otherwise.

Raises

TypeError

The tolerance parameter is not a number.

ValueError

The tolerance parameter is out of [0, 1] range.

Examples using fatf.fairness.models.measures.demographic_parity

Measuring Fairness of a Predictive Model – Disparate Impact