def test_redacting_pd_num(self):
fairness_info = self.creditg_pd_num["fairness_info"]
redacting = Redacting(**fairness_info)
logistic_regression = LogisticRegression(max_iter=1000)
trainable_remi = redacting >> logistic_regression
self._attempt_remi_creditg_pd_num(fairness_info, trainable_remi, 0.80,
0.90)
def fit(self, X, y):
from lale.lib.aif360 import ProtectedAttributesEncoder, Redacting
fairness_info = {
"favorable_labels": self.favorable_labels,
"protected_attributes": self.protected_attributes,
}
redacting = Redacting(**fairness_info) if self.redact else lale.lib.lale.NoOp
trainable_redact_and_prep = redacting >> self.preparation
assert isinstance(trainable_redact_and_prep, lale.operators.TrainablePipeline)
self.redact_and_prep = trainable_redact_and_prep.fit(X, y)
self.prot_attr_enc = ProtectedAttributesEncoder(
**fairness_info,
remainder="drop",
return_X_y=True,
)
prot_attr_names = [pa["feature"] for pa in self.protected_attributes]
self.pandas_to_dataset = _PandasToDatasetConverter(
favorable_label=1,
unfavorable_label=0,
protected_attribute_names=prot_attr_names,
)
encoded_data = self._prep_and_encode(X, y)
self.mitigator.fit(encoded_data)
self.unfavorable_labels = list(set(list(y)) - set(list(self.favorable_labels)))
return self
def fit(self, X, y):
from lale.lib.aif360 import ProtectedAttributesEncoder, Redacting
fairness_info = {
"favorable_labels": self.favorable_labels,
"protected_attributes": self.protected_attributes,
"unfavorable_labels": self.unfavorable_labels,
}
redacting = Redacting(
**fairness_info) if self.redact else lale.lib.lale.NoOp
trainable_redact_and_estim = redacting >> self.estimator
assert isinstance(trainable_redact_and_estim, TrainablePipeline)
self.redact_and_estim = trainable_redact_and_estim.fit(X, y)
self.prot_attr_enc = ProtectedAttributesEncoder(
**fairness_info,
remainder="drop",
return_X_y=True,
)
prot_attr_names = [pa["feature"] for pa in self.protected_attributes]
self.pandas_to_dataset = _PandasToDatasetConverter(
favorable_label=1,
unfavorable_label=0,
protected_attribute_names=prot_attr_names,
)
encoded_X, encoded_y = self.prot_attr_enc.transform(X, y)
self.y_dtype = encoded_y.dtype
self.y_name = encoded_y.name
predicted_y = self.redact_and_estim.predict(X)
predicted_y = _ndarray_to_series(predicted_y, self.y_name, X.index)
_, predicted_y = self.prot_attr_enc.transform(X, predicted_y)
predicted_probas = self.redact_and_estim.predict_proba(X)
dataset_true = self.pandas_to_dataset.convert(encoded_X, encoded_y)
dataset_pred = self.pandas_to_dataset.convert(encoded_X, predicted_y,
predicted_probas)
self.mitigator = self.mitigator.fit(dataset_true, dataset_pred)
self.classes_ = set(list(y))
self.not_favorable_labels = list(self.classes_ -
set(list(self.favorable_labels)))
self.classes_ = np.array(list(self.classes_))
return self
def test_redacting_pd_cat(self):
fairness_info = self.creditg_pd_cat["fairness_info"]
estim = self.prep_pd_cat >> LogisticRegression(max_iter=1000)
trainable_remi = Redacting(**fairness_info) >> estim
self._attempt_remi_creditg_pd_cat(fairness_info, trainable_remi, 0.81, 0.91)