def predict(self, dataset):
"""Perturb the predicted labels to obtain new labels that satisfy
equalized odds constraints.
Args:
dataset (BinaryLabelDataset): Dataset containing labels that needs
to be transformed.
dataset (BinaryLabelDataset): Transformed dataset.
"""
if self.seed is not None:
np.random.seed(self.seed)
# Get the model parameters output from fit
sp2p, sn2p, op2p, on2p = self.model_params.x
# Create boolean conditioning vectors for protected groups
cond_vec_priv = utils.compute_boolean_conditioning_vector(
dataset.protected_attributes, dataset.protected_attribute_names,
self.privileged_groups)
cond_vec_unpriv = utils.compute_boolean_conditioning_vector(
dataset.protected_attributes, dataset.protected_attribute_names,
self.unprivileged_groups)
# Randomly flip labels according to the probabilities in model_params
self_fair_pred = dataset.labels[cond_vec_priv].copy()
self_pp_indices, _ = np.nonzero(
dataset.labels[cond_vec_priv] == dataset.favorable_label)
self_pn_indices, _ = np.nonzero(
dataset.labels[cond_vec_priv] == dataset.unfavorable_label)
np.random.shuffle(self_pp_indices)
np.random.shuffle(self_pn_indices)
n2p_indices = self_pn_indices[:int(len(self_pn_indices) * sn2p)]
self_fair_pred[n2p_indices] = dataset.favorable_label
p2n_indices = self_pp_indices[:int(len(self_pp_indices) * (1 - sp2p))]
self_fair_pred[p2n_indices] = dataset.unfavorable_label
othr_fair_pred = dataset.labels[cond_vec_unpriv].copy()
othr_pp_indices, _ = np.nonzero(
dataset.labels[cond_vec_unpriv] == dataset.favorable_label)
othr_pn_indices, _ = np.nonzero(
dataset.labels[cond_vec_unpriv] == dataset.unfavorable_label)
np.random.shuffle(othr_pp_indices)
np.random.shuffle(othr_pn_indices)
n2p_indices = othr_pn_indices[:int(len(othr_pn_indices) * on2p)]
othr_fair_pred[n2p_indices] = dataset.favorable_label
p2n_indices = othr_pp_indices[:int(len(othr_pp_indices) * (1 - op2p))]
othr_fair_pred[p2n_indices] = dataset.unfavorable_label
# Mutated, fairer dataset with new labels
dataset_new = dataset.copy()
new_labels = np.zeros_like(dataset.labels, dtype=np.float64)
new_labels[cond_vec_priv] = self_fair_pred
new_labels[cond_vec_unpriv] = othr_fair_pred
dataset_new.labels = new_labels
return dataset_new
def __init__(self,
dataset,
unprivileged_groups=None,
privileged_groups=None):
"""
Args:
dataset (StructuredDataset): A StructuredDataset.
privileged_groups (list(dict)): Privileged groups. Format is a list
of `dicts` where the keys are `protected_attribute_names` and
the values are values in `protected_attributes`. Each `dict`
element describes a single group. See examples for more details.
unprivileged_groups (list(dict)): Unprivileged groups in the same
format as `privileged_groups`.
Raises:
TypeError: `dataset` must be a
:obj:`~aiflearn.datasets.StructuredDataset` type.
ValueError: `privileged_groups` and `unprivileged_groups` must be
disjoint.
Examples:
>>> from aiflearn.datasets import GermanDataset
>>> german = GermanDataset()
>>> u = [{'sex': 1, 'age': 1}, {'sex': 0}]
>>> p = [{'sex': 1, 'age': 0}]
>>> dm = DatasetMetric(german, unprivileged_groups=u, privileged_groups=p)
"""
if not isinstance(dataset, StructuredDataset):
raise TypeError("'dataset' should be a StructuredDataset")
# sets self.dataset
super(DatasetMetric, self).__init__(dataset)
# TODO: should this deepcopy?
self.privileged_groups = privileged_groups
self.unprivileged_groups = unprivileged_groups
# don't check if nothing was provided
if not self.privileged_groups or not self.unprivileged_groups:
return
priv_mask = utils.compute_boolean_conditioning_vector(
self.dataset.protected_attributes,
self.dataset.protected_attribute_names, self.privileged_groups)
unpriv_mask = utils.compute_boolean_conditioning_vector(
self.dataset.protected_attributes,
self.dataset.protected_attribute_names, self.unprivileged_groups)
if np.any(np.logical_and(priv_mask, unpriv_mask)):
raise ValueError("'privileged_groups' and 'unprivileged_groups'"
" must be disjoint.")
if not np.all(np.logical_or(priv_mask, unpriv_mask)):
warning("There are some instances in the dataset which are not "
"designated as either privileged or unprivileged. Are you "
"sure this is right?")
def _between_group_generalized_entropy_index(self, groups, alpha=2):
r"""Between-group generalized entropy index is proposed as a group
fairness measure in [2]_ and is one of two terms that the generalized
entropy index decomposes to.
Args:
groups (list): A list of groups over which to calculate this metric.
Groups should be disjoint. By default, this will use the
`privileged_groups` and `unprivileged_groups` as the only two
groups.
alpha (int): See :meth:`generalized_entropy_index`.
References:
.. [2] T. Speicher, H. Heidari, N. Grgic-Hlaca, K. P. Gummadi, A. Singla, A. Weller, and M. B. Zafar,
"A Unified Approach to Quantifying Algorithmic Unfairness: Measuring Individual and Group Unfairness via Inequality Indices,"
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2018.
"""
b = np.zeros(self.dataset.labels.size, dtype=np.float64)
for group in groups:
classified_group = utils.compute_boolean_conditioning_vector(
self.classified_dataset.protected_attributes,
self.classified_dataset.protected_attribute_names,
condition=group)
true_group = utils.compute_boolean_conditioning_vector(
self.dataset.protected_attributes,
self.dataset.protected_attribute_names,
condition=group)
# ignore if there are no members of this group present
if not np.any(true_group):
continue
y_pred = self.classified_dataset.labels[classified_group].ravel()
y_true = self.dataset.labels[true_group].ravel()
y_pred = (
y_pred == self.classified_dataset.favorable_label).astype(
np.float64)
y_true = (y_true == self.dataset.favorable_label).astype(
np.float64)
b[true_group] = np.mean(1 + y_pred - y_true)
if alpha == 1:
return np.mean(np.log((b / np.mean(b))**b) / np.mean(b))
elif alpha == 0:
return -np.mean(np.log(b / np.mean(b)) / np.mean(b))
else:
return np.mean((b / np.mean(b))**alpha - 1) / (alpha * (alpha - 1))
def predict(self, dataset):
"""Obtain fair predictions using the ROC method.
Args:
dataset (BinaryLabelDataset): Dataset containing scores that will
be used to compute predicted labels.
Returns:
dataset_pred (BinaryLabelDataset): Output dataset with potentially
fair predictions obtain using the ROC method.
"""
dataset_new = dataset.copy(deepcopy=False)
fav_pred_inds = (dataset.scores > self.classification_threshold)
unfav_pred_inds = ~fav_pred_inds
y_pred = np.zeros(dataset.scores.shape)
y_pred[fav_pred_inds] = dataset.favorable_label
y_pred[unfav_pred_inds] = dataset.unfavorable_label
# Indices of critical region around the classification boundary
crit_region_inds = np.logical_and(
dataset.scores <= self.classification_threshold + self.ROC_margin,
dataset.scores > self.classification_threshold - self.ROC_margin)
# Indices of privileged and unprivileged groups
cond_priv = utils.compute_boolean_conditioning_vector(
dataset.protected_attributes, dataset.protected_attribute_names,
self.privileged_groups)
cond_unpriv = utils.compute_boolean_conditioning_vector(
dataset.protected_attributes, dataset.protected_attribute_names,
self.unprivileged_groups)
# New, fairer labels
dataset_new.labels = y_pred
dataset_new.labels[np.logical_and(crit_region_inds,
cond_priv.reshape(
-1,
1))] = dataset.unfavorable_label
dataset_new.labels[np.logical_and(crit_region_inds,
cond_unpriv.reshape(
-1,
1))] = dataset.favorable_label
return dataset_new
def predict(self, dataset):
"""Perturb the predicted scores to obtain new labels that satisfy
equalized odds constraints, while preserving calibration.
Args:
dataset (BinaryLabelDataset): Dataset containing `scores` that needs
to be transformed.
Returns:
dataset (BinaryLabelDataset): transformed dataset.
"""
if self.seed is not None:
np.random.seed(self.seed)
cond_vec_priv = utils.compute_boolean_conditioning_vector(
dataset.protected_attributes, dataset.protected_attribute_names,
self.privileged_groups)
cond_vec_unpriv = utils.compute_boolean_conditioning_vector(
dataset.protected_attributes, dataset.protected_attribute_names,
self.unprivileged_groups)
priv_indices = (np.random.random(sum(cond_vec_priv)) <=
self.priv_mix_rate)
priv_new_pred = dataset.scores[cond_vec_priv].copy()
priv_new_pred[priv_indices] = self.base_rate_priv
unpriv_indices = (np.random.random(sum(cond_vec_unpriv)) <=
self.unpriv_mix_rate)
unpriv_new_pred = dataset.scores[cond_vec_unpriv].copy()
unpriv_new_pred[unpriv_indices] = self.base_rate_unpriv
dataset_new = dataset.copy(deepcopy=True)
dataset_new.scores = np.zeros_like(dataset.scores, dtype=np.float64)
dataset_new.scores[cond_vec_priv] = priv_new_pred
dataset_new.scores[cond_vec_unpriv] = unpriv_new_pred
# Create labels from scores using a default threshold
dataset_new.labels = np.where(dataset_new.scores >= self.threshold,
dataset_new.favorable_label,
dataset_new.unfavorable_label)
return dataset_new
def _obtain_conditionings(self, dataset):
"""Obtain the necessary conditioning boolean vectors to compute
instance level weights.
"""
# conditioning
priv_cond = utils.compute_boolean_conditioning_vector(
dataset.protected_attributes,
dataset.protected_attribute_names,
condition=self.privileged_groups)
unpriv_cond = utils.compute_boolean_conditioning_vector(
dataset.protected_attributes,
dataset.protected_attribute_names,
condition=self.unprivileged_groups)
fav_cond = dataset.labels.ravel() == dataset.favorable_label
unfav_cond = dataset.labels.ravel() == dataset.unfavorable_label
# combination of label and privileged/unpriv. groups
cond_p_fav = np.logical_and(fav_cond, priv_cond)
cond_p_unfav = np.logical_and(unfav_cond, priv_cond)
cond_up_fav = np.logical_and(fav_cond, unpriv_cond)
cond_up_unfav = np.logical_and(unfav_cond, unpriv_cond)
return (priv_cond, unpriv_cond, fav_cond, unfav_cond, cond_p_fav,
cond_p_unfav, cond_up_fav, cond_up_unfav)
def fit(self, dataset_true, dataset_pred):
"""Compute parameters for equalizing odds using true and predicted
labels.
Args:
true_dataset (BinaryLabelDataset): Dataset containing true labels.
pred_dataset (BinaryLabelDataset): Dataset containing predicted
labels.
Returns:
EqOddsPostprocessing: Returns self.
"""
metric = ClassificationMetric(
dataset_true,
dataset_pred,
unprivileged_groups=self.unprivileged_groups,
privileged_groups=self.privileged_groups)
# compute basic statistics
sbr = metric.num_instances(privileged=True) / metric.num_instances()
obr = metric.num_instances(privileged=False) / metric.num_instances()
fpr0 = metric.false_positive_rate(privileged=True)
fpr1 = metric.false_positive_rate(privileged=False)
fnr0 = metric.false_negative_rate(privileged=True)
fnr1 = metric.false_negative_rate(privileged=False)
tpr0 = metric.true_positive_rate(privileged=True)
tpr1 = metric.true_positive_rate(privileged=False)
tnr0 = metric.true_negative_rate(privileged=True)
tnr1 = metric.true_negative_rate(privileged=False)
# linear program has 4 decision variables:
# [Pr[label_tilde = 1 | label_hat = 1, protected_attributes = 0];
# Pr[label_tilde = 1 | label_hat = 0, protected_attributes = 0];
# Pr[label_tilde = 1 | label_hat = 1, protected_attributes = 1];
# Pr[label_tilde = 1 | label_hat = 0, protected_attributes = 1]]
# Coefficients of the linear objective function to be minimized.
c = np.array([fpr0 - tpr0, tnr0 - fnr0, fpr1 - tpr1, tnr1 - fnr1])
# A_ub - 2-D array which, when matrix-multiplied by x, gives the values
# of the upper-bound inequality constraints at x
# b_ub - 1-D array of values representing the upper-bound of each
# inequality constraint (row) in A_ub.
# Just to keep these between zero and one
A_ub = np.array(
[[1, 0, 0, 0], [-1, 0, 0, 0], [0, 1, 0, 0], [0, -1, 0, 0],
[0, 0, 1, 0], [0, 0, -1, 0], [0, 0, 0, 1], [0, 0, 0, -1]],
dtype=np.float64)
b_ub = np.array([1, 0, 1, 0, 1, 0, 1, 0], dtype=np.float64)
# Create boolean conditioning vectors for protected groups
cond_vec_priv = utils.compute_boolean_conditioning_vector(
dataset_pred.protected_attributes,
dataset_pred.protected_attribute_names, self.privileged_groups)
cond_vec_unpriv = utils.compute_boolean_conditioning_vector(
dataset_pred.protected_attributes,
dataset_pred.protected_attribute_names, self.unprivileged_groups)
sconst = np.ravel(
dataset_pred.labels[cond_vec_priv] == dataset_pred.favorable_label)
sflip = np.ravel(dataset_pred.labels[cond_vec_priv] ==
dataset_pred.unfavorable_label)
oconst = np.ravel(dataset_pred.labels[cond_vec_unpriv] ==
dataset_pred.favorable_label)
oflip = np.ravel(dataset_pred.labels[cond_vec_unpriv] ==
dataset_pred.unfavorable_label)
y_true = dataset_true.labels.ravel()
sm_tn = np.logical_and(
sflip,
y_true[cond_vec_priv] == dataset_true.unfavorable_label,
dtype=np.float64)
sm_fn = np.logical_and(
sflip,
y_true[cond_vec_priv] == dataset_true.favorable_label,
dtype=np.float64)
sm_fp = np.logical_and(
sconst,
y_true[cond_vec_priv] == dataset_true.unfavorable_label,
dtype=np.float64)
sm_tp = np.logical_and(
sconst,
y_true[cond_vec_priv] == dataset_true.favorable_label,
dtype=np.float64)
om_tn = np.logical_and(
oflip,
y_true[cond_vec_unpriv] == dataset_true.unfavorable_label,
dtype=np.float64)
om_fn = np.logical_and(
oflip,
y_true[cond_vec_unpriv] == dataset_true.favorable_label,
dtype=np.float64)
om_fp = np.logical_and(
oconst,
y_true[cond_vec_unpriv] == dataset_true.unfavorable_label,
dtype=np.float64)
om_tp = np.logical_and(
oconst,
y_true[cond_vec_unpriv] == dataset_true.favorable_label,
dtype=np.float64)
# A_eq - 2-D array which, when matrix-multiplied by x,
# gives the values of the equality constraints at x
# b_eq - 1-D array of values representing the RHS of each equality
# constraint (row) in A_eq.
# Used to impose equality of odds constraint
A_eq = [
[(np.mean(sconst * sm_tp) - np.mean(sflip * sm_tp)) / sbr,
(np.mean(sflip * sm_fn) - np.mean(sconst * sm_fn)) / sbr,
(np.mean(oflip * om_tp) - np.mean(oconst * om_tp)) / obr,
(np.mean(oconst * om_fn) - np.mean(oflip * om_fn)) / obr],
[(np.mean(sconst * sm_fp) - np.mean(sflip * sm_fp)) / (1 - sbr),
(np.mean(sflip * sm_tn) - np.mean(sconst * sm_tn)) / (1 - sbr),
(np.mean(oflip * om_fp) - np.mean(oconst * om_fp)) / (1 - obr),
(np.mean(oconst * om_tn) - np.mean(oflip * om_tn)) / (1 - obr)]
]
b_eq = [
(np.mean(oflip * om_tp) + np.mean(oconst * om_fn)) / obr -
(np.mean(sflip * sm_tp) + np.mean(sconst * sm_fn)) / sbr,
(np.mean(oflip * om_fp) + np.mean(oconst * om_tn)) / (1 - obr) -
(np.mean(sflip * sm_fp) + np.mean(sconst * sm_tn)) / (1 - sbr)
]
# Linear program
self.model_params = linprog(c,
A_ub=A_ub,
b_ub=b_ub,
A_eq=A_eq,
b_eq=b_eq)
return self
def fit(self, dataset_true, dataset_pred):
"""Compute parameters for equalizing generalized odds using true and
predicted scores, while preserving calibration.
Args:
dataset_true (BinaryLabelDataset): Dataset containing true `labels`.
dataset_pred (BinaryLabelDataset): Dataset containing predicted
`scores`.
Returns:
CalibratedEqOddsPostprocessing: Returns self.
"""
# Create boolean conditioning vectors for protected groups
cond_vec_priv = utils.compute_boolean_conditioning_vector(
dataset_pred.protected_attributes,
dataset_pred.protected_attribute_names, self.privileged_groups)
cond_vec_unpriv = utils.compute_boolean_conditioning_vector(
dataset_pred.protected_attributes,
dataset_pred.protected_attribute_names, self.unprivileged_groups)
cm = ClassificationMetric(dataset_true,
dataset_pred,
unprivileged_groups=self.unprivileged_groups,
privileged_groups=self.privileged_groups)
self.base_rate_priv = cm.base_rate(privileged=True)
self.base_rate_unpriv = cm.base_rate(privileged=False)
# Create a dataset with "trivial" predictions
dataset_trivial = dataset_pred.copy(deepcopy=True)
dataset_trivial.scores[cond_vec_priv] = cm.base_rate(privileged=True)
dataset_trivial.scores[cond_vec_unpriv] = cm.base_rate(
privileged=False)
cm_triv = ClassificationMetric(
dataset_true,
dataset_trivial,
unprivileged_groups=self.unprivileged_groups,
privileged_groups=self.privileged_groups)
if self.fn_rate == 0:
priv_cost = cm.generalized_false_positive_rate(privileged=True)
unpriv_cost = cm.generalized_false_positive_rate(privileged=False)
priv_trivial_cost = cm_triv.generalized_false_positive_rate(
privileged=True)
unpriv_trivial_cost = cm_triv.generalized_false_positive_rate(
privileged=False)
elif self.fp_rate == 0:
priv_cost = cm.generalized_false_negative_rate(privileged=True)
unpriv_cost = cm.generalized_false_negative_rate(privileged=False)
priv_trivial_cost = cm_triv.generalized_false_negative_rate(
privileged=True)
unpriv_trivial_cost = cm_triv.generalized_false_negative_rate(
privileged=False)
else:
priv_cost = weighted_cost(self.fp_rate,
self.fn_rate,
cm,
privileged=True)
unpriv_cost = weighted_cost(self.fp_rate,
self.fn_rate,
cm,
privileged=False)
priv_trivial_cost = weighted_cost(self.fp_rate,
self.fn_rate,
cm_triv,
privileged=True)
unpriv_trivial_cost = weighted_cost(self.fp_rate,
self.fn_rate,
cm_triv,
privileged=False)
unpriv_costs_more = unpriv_cost > priv_cost
self.priv_mix_rate = (unpriv_cost - priv_cost) / (
priv_trivial_cost - priv_cost) if unpriv_costs_more else 0
self.unpriv_mix_rate = 0 if unpriv_costs_more else (
priv_cost - unpriv_cost) / (unpriv_trivial_cost - unpriv_cost)
return self