def test_adv_debias_old_reproduce():
"""Test that the old AdversarialDebiasing is reproducible."""
sess = tf.Session()
old_adv_deb = OldAdversarialDebiasing(unprivileged_groups=[{
'sex': 0
}],
privileged_groups=[{
'sex': 1
}],
scope_name='old_classifier',
sess=sess,
num_epochs=5,
seed=123)
old_preds = old_adv_deb.fit_predict(adult)
sess.close()
tf.reset_default_graph()
sess = tf.Session()
old_adv_deb2 = OldAdversarialDebiasing(unprivileged_groups=[{
'sex': 0
}],
privileged_groups=[{
'sex': 1
}],
scope_name='old_classifier',
sess=sess,
num_epochs=5,
seed=123)
old_preds2 = old_adv_deb2.fit_predict(adult)
sess.close()
assert np.allclose(old_preds.labels, old_preds2.labels)
def _preprocess_data(
data, protected_attribute_name, protected_attribute_index, label_name, required_fairness
):
from pandas import DataFrame
from aif360.datasets import BinaryLabelDataset
dataset = BinaryLabelDataset(
df=DataFrame(data),
protected_attribute_names={protected_attribute_name},
label_names={label_name},
favorable_label=2,
unfavorable_label=1,
)
train, test = dataset.split([0.8])
from aif360.algorithms.inprocessing import AdversarialDebiasing
sess = tf.compat.v1.Session()
debiaser = AdversarialDebiasing(
unprivileged_groups=({protected_attribute_name: 0},),
privileged_groups=({protected_attribute_name: 1},),
scope_name="debiaser",
debias=True,
sess=sess,
)
debiaser.fit(train)
from sklearn.ensemble import RandomForestClassifier
model = RandomForestClassifier(class_weight="balanced")
X_tr = np.delete(train.features, protected_attribute_index, axis=1)
y_tr = train.labels.ravel()
model.fit(X_tr, y_tr)
test_pred = test.copy(deepcopy=True)
test_pred.scores = model.predict(np.delete(debiaser.predict(test).features, protected_attribute_index, axis=1))
accuracy = np.sum(np.equal(test.scores, test_pred.scores))
from aif360.metrics import ClassificationMetric
disparate_impact = ClassificationMetric(
test,
test_pred,
unprivileged_groups=({protected_attribute_name: 0},),
privileged_groups=({protected_attribute_name: 1},),
).disparate_impact()
print(f"Accuracy: {accuracy}")
print(f"Disparate impact: {disparate_impact}")
if disparate_impact > float(required_fairness):
raise ValueError(
f"Too unfair! Disparate impact was {disparate_impact} but must be less than {required_fairness}"
)
def nondebiased_classifier(train, test, privileged_groups,
unprivileged_groups):
sess = tf.Session()
NN_model = AdversarialDebiasing(privileged_groups,
unprivileged_groups,
scope_name='nondebiased_classifier',
debias=False,
sess=sess)
NN_model.fit(train)
# predict outcome using the test set
pred_NNmodel = NN_model.predict(test)
sess.close()
tf.reset_default_graph()
# calculate accuracy
accuracy = accuracy_score(y_true=test.labels, y_pred=pred_NNmodel.labels)
# calculate fairness metrics
metric_test = BinaryLabelDatasetMetric(
pred_NNmodel,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
acc_test = ClassificationMetric(test,
pred_NNmodel,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
equal_opportunity_difference = equal_opp_diff(test,
pred_NNmodel,
'sex',
privileged=1,
unprivileged=0,
favourable=1,
unfavourable=0)
average_odds_difference = avg_odds_diff(test,
pred_NNmodel,
'sex',
privileged=1,
unprivileged=0,
favourable=1,
unfavourable=0)
metrics = [
metric_test.mean_difference(),
acc_test.disparate_impact(), equal_opportunity_difference,
average_odds_difference,
acc_test.theil_index()
]
return pred_NNmodel, accuracy, metrics
def __init__(self, df, target_col, sensitive_att, seed=0):
"""
:param df: pandas dataframe, stores the data to fit the fair classifier.
:param target_col: str, the name of the target variable in above data.
:param sensitive_att: str, the name of a sensitive attribute in above data. If none, call auto_detection to update. Value 0 represent protected.
:param seed: integer, random seed.
"""
import tensorflow as tf
sess = tf.Session()
cur_step = AdversarialDebiasing(unprivileged_groups=[{
sensitive_att: 0
}],
privileged_groups=[{
sensitive_att: 1
}],
scope_name='debiased_classifier',
debias=True,
sess=sess,
seed=seed)
super().__init__("@".join(["AIF_AdversarialDebiasing", sensitive_att]),
cur_step,
df,
target_col,
sensitive_att=sensitive_att,
fair_aware=True)
def test_adv_debias_old():
"""Test that the predictions of the old and new AdversarialDebiasing match.
"""
tf.reset_default_graph()
sess = tf.Session()
old_adv_deb = OldAdversarialDebiasing(unprivileged_groups=[{
'sex': 0
}],
privileged_groups=[{
'sex': 1
}],
scope_name='old_classifier',
sess=sess,
num_epochs=5,
seed=123)
old_preds = old_adv_deb.fit_predict(adult)
sess.close()
adv_deb = AdversarialDebiasing('sex', num_epochs=5, random_state=123)
new_preds = adv_deb.fit(X, y).predict(X)
adv_deb.sess_.close()
assert np.allclose(old_preds.labels.flatten(), new_preds)
def adversarial_debias(structured_data_train,
structured_data_test,
priv_category,
scope_name='debiased_classifier',
num_epochs=10):
"""
Remove bias from dataset using Adversarial debiasing. Must use Tensorflow version 1.14.
Parameters:
structured_data (aif360.datasets.standard_dataset.StandardDataset): Structured dataset.
priv_category (string): Column with privileged class.
scope_name (string): Name of scope to debias. See documentation on adversarial debiasing for details.
num_epochs (int): Epochs to train with.
Returns:
data_transf_train_df (pandas dataframe): Pandas dataframe of train set
data_transf_df (pandas dataframe): Pandas dataframe of test set
"""
sess = tf.Session()
# Get privileged and unprivileged groups
privileged_groups, unprivileged_groups = get_attributes(
structured_data_train, selected_attr=[priv_category])
debiased_model = AdversarialDebiasing(
privileged_groups=privileged_groups,
unprivileged_groups=unprivileged_groups,
scope_name=scope_name,
num_epochs=num_epochs,
debias=True,
sess=sess)
debiased_model.fit(structured_data_train)
# Remove bias
data_train_pred = debiased_model.predict(structured_data_train)
data_pred = debiased_model.predict(structured_data_test)
# Convert to pandas dataframe
data_transf_train_df = convert_to_pd_dataframe(data_train_pred)
data_transf_test_df = convert_to_pd_dataframe(data_pred)
return data_transf_train_df, data_transf_test_df