aif360.sklearn.postprocessing.RejectOptionClassifier

class aif360.sklearn.postprocessing.RejectOptionClassifier(prot_attr=None, threshold=0.5, margin=0.1)

Reject option based classification (ROC) post-processor.

Reject option classification is a post-processing technique that gives favorable outcomes to unprivileged groups and unfavorable outcomes to privileged groups in a confidence band around the decision boundary with the highest uncertainty [1].

Note

A Pipeline expects a single estimation step but this class requires an estimator’s predictions as input. See PostProcessingMeta for a workaround.

See also

PostProcessingMeta, RejectOptionClassifierCV

References

[1]	F. Kamiran, A. Karim, and X. Zhang, “Decision Theory for Discrimination-Aware Classification,” IEEE International Conference on Data Mining, 2012.

Variables:

• prot_attr (str or list(str)) – Protected attribute(s) used for post- processing.
• groups (array, shape (2,)) – A list of group labels known to the classifier. Note: this algorithm require a binary division of the data.
• classes (array, shape (num_classes,)) – A list of class labels known to the classifier. Note: this algorithm treats all non-positive outcomes as negative (binary classification only).
• pos_label (scalar) – The label of the positive class.
• priv_group (scalar) – The label of the privileged group.

Examples

RejectOptionClassifier can be easily paired with GridSearchCV to find the best threshold and margin with respect to a fairness measure:

>>> from sklearn.model_selection import GridSearchCV
>>> roc = RejectOptionClassifier()
>>> param = [{'threshold': [t],
              'margin': np.arange(0.05, min(t, 1-t)+0.025, 0.05)}
...          for t in np.arange(0.05, 1., 0.05)]
>>> stat_par = make_scorer(statistical_parity_difference)
>>> scoring = {'bal_acc': 'balanced_accuracy', 'stat_par': stat_par}
>>> def refit(cv_res):
...     return np.ma.array(cv_res['mean_test_bal_acc'],
...             mask=cv_res['mean_test_stat_par'] < -0.1).argmax()
...
>>> grid = GridSearchCV(roc, param, scoring=scoring, refit=refit)

Or, alternatively, this can be done in one step with RejectOptionClassifierCV:

>>> grid = RejectOptionClassifierCV(scoring='statistical_parity')

Parameters:

• prot_attr (single label or list-like, optional) – Protected attribute(s) to use in the post-processing. If more than one attribute, all combinations of values (intersections) are considered. Default is None meaning all protected attributes from the dataset are used. Note: This algorithm requires there be exactly 2 groups (privileged and unprivileged).
• threshold (scalar) – Classification threshold. Probability estimates greater than this value are considered positive. Must be between 0 and 1.
• margin (scalar) – Half width of the critical region. Estimates within the critical region are “rejected” and assigned according to their group. Must be between 0 and min(threshold, 1-threshold).
• ('statistical_parity', 'average_odds', 'equal_opportunity', (metric) – or callable):

Methods

fit	This is essentially a no-op; it simply validates the inputs and stores them for predict.
fit_predict	Predict class labels for the given scores.
get_params	Get parameters for this estimator.
predict	Predict class labels for the given scores.
predict_proba	Probability estimates.
score	Return the mean accuracy on the given test data and labels.
set_params	Set the parameters of this estimator.

__init__(prot_attr=None, threshold=0.5, margin=0.1)

Parameters:

• prot_attr (single label or list-like, optional) – Protected attribute(s) to use in the post-processing. If more than one attribute, all combinations of values (intersections) are considered. Default is None meaning all protected attributes from the dataset are used. Note: This algorithm requires there be exactly 2 groups (privileged and unprivileged).
• threshold (scalar) – Classification threshold. Probability estimates greater than this value are considered positive. Must be between 0 and 1.
• margin (scalar) – Half width of the critical region. Estimates within the critical region are “rejected” and assigned according to their group. Must be between 0 and min(threshold, 1-threshold).
• ('statistical_parity', 'average_odds', 'equal_opportunity', (metric) – or callable):

fit(X, y, labels=None, pos_label=1, priv_group=1, sample_weight=None)

This is essentially a no-op; it simply validates the inputs and stores them for predict.

Parameters:

• X (array-like) – Ignored.
• y (array-like) – Ground-truth (correct) target values. Note: one of X or y must contain protected attribute information.
• labels (list, optional) – The ordered set of labels values. Must match the order of columns in X if provided. By default, all labels in y are used in sorted order.
• pos_label (scalar, optional) – The label of the positive class.
• priv_group (scalar, optional) – The label of the privileged group.
• sample_weight (array-like, optional) – Ignored.

Returns:

self

fit_predict(X, y=None, **fit_params)

Predict class labels for the given scores.

In general, it is not necessary to fit and predict separately so this method may be used instead. For subsequent predicts, it may be easier to use the predict method, though.

Parameters:

• X (pandas.DataFrame) – Probability estimates of the targets as returned by a predict_proba() call or equivalent. Note: must include protected attributes in the index.
• y (array-like, optional) – Ground-truth (correct) target values. Note: if not provided, labels must be provided in **fit_params. See fit for details.
• **fit_params – See fit for details.

Returns:

numpy.ndarray – Predicted class label per sample.

predict(X)

Predict class labels for the given scores.

Parameters:X (pandas.DataFrame) – Probability estimates of the targets as returned by a predict_proba() call or equivalent. Note: must include protected attributes in the index. Returns:numpy.ndarray – Predicted class label per sample.

predict_proba(X)

Probability estimates.

The returned estimates for all classes are ordered by the label of classes.

Parameters:X (pandas.DataFrame) – Probability estimates of the targets as returned by a predict_proba() call or equivalent. Note: must include protected attributes in the index. Returns:numpy.ndarray – Returns the probability of the sample for each class in the model, where classes are ordered as they are in self.classes_.