class Fair_PR():
def __init__(self, sensitive, class_attr, eta):
self.model = PrejudiceRemover(sensitive_attr=sensitive,
class_attr=class_attr,
eta=eta)
def fit(self, data, labels, prot):
ds = BinaryLabelDataset(df=data,
label_names=labels,
protected_attribute_names=prot)
self.model.fit(ds)
def predict_proba(self, data_test):
y = self.model.predict(data_test)
return y.scores
def prejudice_remover(dataset_orig_train, dataset_orig_valid,
dataset_orig_test, privileged_groups,
unprivileged_groups, sens_attr):
#print(dataset_orig_train.features.shape)
#print(dataset_orig_train.labels.shape )
dataset_original_train = copy.deepcopy(dataset_orig_train)
dataset_original_valid = copy.deepcopy(dataset_orig_valid)
dataset_original_test = copy.deepcopy(dataset_orig_test)
# Learning a Prejudice Remover (PR) model on original data
model = PrejudiceRemover(sensitive_attr=sens_attr, eta=25.0)
pr_orig_scaler = StandardScaler()
dataset_original_train.features = pr_orig_scaler.fit_transform(
dataset_original_train.features)
dataset_original_valid.features = pr_orig_scaler.fit_transform(
dataset_original_valid.features)
dataset_original_test.features = pr_orig_scaler.fit_transform(
dataset_original_test.features)
pr_orig = model.fit(dataset_original_train)
# validating PR model
thresh_arr = np.linspace(0.01, 0.50, 50)
valid_pred, valid_metrics = test(dataset_original_valid, pr_orig,
thresh_arr, privileged_groups,
unprivileged_groups)
pr_orig_best_ind = np.argmax(valid_metrics['bal_acc'])
train_pred, metric_train = test(dataset_original_train, pr_orig,
[thresh_arr[pr_orig_best_ind]],
privileged_groups, unprivileged_groups)
val_pred, metric_val = test(dataset_original_valid, pr_orig,
[thresh_arr[pr_orig_best_ind]],
privileged_groups, unprivileged_groups)
test_pred, metric_test = test(dataset_original_test, pr_orig,
[thresh_arr[pr_orig_best_ind]],
privileged_groups, unprivileged_groups)
dataset_transf_valid_pred = copy.deepcopy(dataset_orig_valid)
dataset_transf_valid_pred.labels = val_pred.labels
print(val_pred.labels.shape)
#dataset_transf_valid_pred.scores = val_pred.scores
dataset_transf_test_pred = copy.deepcopy(dataset_orig_test)
dataset_transf_test_pred.labels = test_pred.labels
#dataset_transf_test_pred.scores = test_pred.scores
return dataset_transf_valid_pred, dataset_transf_test_pred
def prejudice(train, test, unprivileged_groups, privileged_groups):
prejudice_model = PrejudiceRemover(eta=100, sensitive_attr='sex')
prejudice_model.fit(train)
# predict outcome using the test set
pred_prejudice = prejudice_model.predict(test)
# calculate accuracy
accuracy = accuracy_score(y_true=test.labels, y_pred=pred_prejudice.labels)
# calculate fairness metrics
metric_test = BinaryLabelDatasetMetric(
pred_prejudice,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
acc_test = ClassificationMetric(test,
pred_prejudice,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
equal_opportunity_difference = equal_opp_diff(test,
pred_prejudice,
'sex',
privileged=1,
unprivileged=0,
favourable=1,
unfavourable=0)
average_odds_difference = avg_odds_diff(test,
pred_prejudice,
'sex',
privileged=1,
unprivileged=0,
favourable=1,
unfavourable=0)
if acc_test.disparate_impact() == math.inf:
disparate_impact = 5.0
else:
disparate_impact = acc_test.disparate_impact()
metrics = [
metric_test.mean_difference(), disparate_impact,
equal_opportunity_difference, average_odds_difference,
acc_test.theil_index()
]
return pred_prejudice, accuracy, metrics
def prejudice_remover(structured_data_train,
structured_data_test,
priv_category,
eta=25):
"""
Remove bias from dataset using Prejudice Remover Regularizer.
Parameters:
structured_data (aif360.datasets.standard_dataset.StandardDataset): Structured dataset.
priv_category (string): Column with privileged class.
eta (int): Regularization parameter
Returns:
data_transf_train_df (pandas dataframe): Pandas dataframe.
data_transf_test_df (pandas dataframe): Pandas dataframe.
"""
debiased_model = PrejudiceRemover(sensitive_attr=priv_category, eta=eta)
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
def __init__(self, df, target_col, sensitive_att, fairness_penalty=1.0):
"""
: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 fairness_penalty: float in [0,1], fairness penalty parameter. default is 1. The same parameter in aif360.algorithms.inprocessing.PrejudiceRemover.
"""
# TODO: fix the bug that cannot import lib of 'getoutput'
cur_step = PrejudiceRemover(eta=fairness_penalty,
sensitive_attr=sensitive_att,
class_attr=target_col)
super().__init__("@".join(["AIF_PrejudiceRemover", sensitive_att]),
cur_step,
df,
target_col,
sensitive_att=sensitive_att,
fair_aware=True)
def run_trial():
# stores each run (4 algos without reweighing, 4 algos with reweighing) as a sublist
# number of sublists = number of runs
# each sublist has four elements
# we ONLY predict on the testing data
########## WITHOUT REWEIGHING #############
stat_par = []
disp_imp = []
eq_opp_diff = []
avg_odds_diff = []
theil = []
acc = []
########## WITH REWEIGHING #############
stat_par_reweigh = []
disp_imp_reweigh = []
eq_opp_diff_reweigh = []
avg_odds_diff_reweigh = []
theil_reweigh = []
acc_reweigh = []
###########################################
for i in range(10):
###########################################
privileged_groups = [{'sex': 1}]
unprivileged_groups = [{'sex': 0}]
dataset_orig = load_preproc_data_adult()
# train1, test1 are the original dataset
train1, test1 = dataset_orig.split([0.7], shuffle=True)
RW = Reweighing(unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
RW.fit(train1)
# dataset_transf_train, test1 are for the reweighed dataset
dataset_transf_train = RW.transform(train1)
###########################################
# change weights to whole numbers
for i in range(dataset_transf_train.instance_weights.size):
dataset_transf_train.instance_weights[i] = (round(
dataset_transf_train.instance_weights[i] / 0.1) * 0.1) * 10
weights = copy.deepcopy(dataset_transf_train.instance_weights)
# change dataset_transf_train.features and dataset_transf_train.labels and dataset_transf_train.protected_attributes according to the weights of each instance
sum_weights = 0
for i in range(dataset_transf_train.features.shape[0]):
row = copy.deepcopy(dataset_transf_train.features[i])
row_label = copy.deepcopy(dataset_transf_train.labels[i])
row_protected_attributes = copy.deepcopy(
dataset_transf_train.protected_attributes[i])
row_protected_attributes.resize(1, 2)
row.resize(1, 18)
row_label.resize(1, 1)
weight = int(weights[i])
for j in range(weight - 1):
dataset_transf_train.features = np.concatenate(
(dataset_transf_train.features, row))
dataset_transf_train.labels = np.concatenate(
(dataset_transf_train.labels, row_label))
dataset_transf_train.protected_attributes = np.concatenate(
(dataset_transf_train.protected_attributes,
row_protected_attributes))
# change the dataset_transf_train to a numpy array of ones to match number of rows in features
dataset_transf_train.instance_weights = np.ones(
dataset_transf_train.features.shape[0])
################## without reweighing ##########################
temp_stat_par = []
temp_disp_imp = []
temp_eq_opp_diff = []
temp_avg_odds_diff = []
temp_theil = []
temp_acc = []
##################### adversarial debiasing #####################
sess = tf.Session()
debiased_model = AdversarialDebiasing(
privileged_groups=privileged_groups,
unprivileged_groups=unprivileged_groups,
scope_name='debiased_classifier',
debias=True,
sess=sess)
debiased_model.fit(train1)
dataset_debiasing_test = debiased_model.predict(test1)
sess.close()
tf.reset_default_graph()
##################### metrics #####################
metric_test = BinaryLabelDatasetMetric(
dataset_debiasing_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
acc_test = ClassificationMetric(
test1,
dataset_debiasing_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
temp_stat_par.append(metric_test.mean_difference())
temp_disp_imp.append(metric_test.disparate_impact())
temp_eq_opp_diff.append(metric_test.equal_opportunity_difference())
temp_avg_odds_diff.append(metric_test.average_odds_difference())
temp_theil.append(metric_test.theil_index())
temp_acc.append(acc_test.accuracy())
##################### prejudice remover #####################
prejudice_model = PrejudiceRemover(eta=100, sensitive_attr='sex')
prejudice_model.fit(train1)
dataset_prejudice_test = prejudice_model.predict(test1)
##################### metrics #####################
metric_test = BinaryLabelDatasetMetric(
dataset_prejudice_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
acc_test = ClassificationMetric(
test1,
dataset_prejudice_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
temp_stat_par.append(metric_test.mean_difference())
temp_disp_imp.append(metric_test.disparate_impact())
temp_eq_opp_diff.append(metric_test.equal_opportunity_difference())
temp_avg_odds_diff.append(metric_test.average_odds_difference())
temp_theil.append(metric_test.theil_index())
temp_acc.append(acc_test.accuracy())
##################### normal neural net #####################
sess = tf.Session()
neural_model = AdversarialDebiasing(
privileged_groups=privileged_groups,
unprivileged_groups=unprivileged_groups,
scope_name='debiased_classifier',
debias=False,
sess=sess)
neural_model.fit(train1)
dataset_neural_test = neural_model.predict(test1)
sess.close()
tf.reset_default_graph()
##################### metrics #####################
metric_test = BinaryLabelDatasetMetric(
dataset_neural_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
acc_test = ClassificationMetric(
test1,
dataset_neural_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
temp_stat_par.append(metric_test.mean_difference())
temp_disp_imp.append(metric_test.disparate_impact())
temp_eq_opp_diff.append(metric_test.equal_opportunity_difference())
temp_avg_odds_diff.append(metric_test.average_odds_difference())
temp_theil.append(metric_test.theil_index())
temp_acc.append(acc_test.accuracy())
##################### ensemble #####################
pred_labels_test = []
for i in range(0, len(test1.features)):
arr_test = mode([
dataset_debiasing_test[i], dataset_prejudice_test[i],
dataset_neural_test[i]
])
pred_labels_test.append(arr_test[0][0])
dataset_ensemble_test = test1.copy()
dataset_ensemble_test.labels = np.array(pred_labels_test)
##################### metrics #####################
metric_test = BinaryLabelDatasetMetric(
dataset_ensemble_train,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
acc_test = ClassificationMetric(
test1,
dataset_ensemble_train,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
temp_stat_par.append(metric_test.mean_difference())
temp_disp_imp.append(metric_test.disparate_impact())
temp_eq_opp_diff.append(metric_test.equal_opportunity_difference())
temp_avg_odds_diff.append(metric_test.average_odds_difference())
temp_theil.append(metric_test.theil_index())
temp_acc.append(acc_test.accuracy())
######### DUMP SHIT ###########
stat_par.append(temp_stat_par)
disp_imp.append(temp_disp_imp)
eq_opp_diff.append(temp_eq_opp_diff)
avg_odds_diff.append(temp_avg_odds_diff)
theil.append(temp_theil)
acc.append(temp_acc)
################## with reweighing ##########################
temp_stat_par = []
temp_disp_imp = []
temp_eq_opp_diff = []
temp_avg_odds_diff = []
temp_theil = []
temp_acc = []
################## adversarial debiasing ##################
sess = tf.Session()
debiased_model_reweighing = AdversarialDebiasing(
privileged_groups=privileged_groups,
unprivileged_groups=unprivileged_groups,
scope_name='debiased_classifier',
debias=True,
sess=sess)
debiased_model_reweighing.fit(dataset_transf_train)
dataset_debiasing_test_reweighing = debiased_model_reweighing.predict(
test1)
sess.close()
tf.reset_default_graph()
##################### metrics #####################
metric_test = BinaryLabelDatasetMetric(
dataset_debiasing_test_reweighing,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
acc_test = ClassificationMetric(
test1,
dataset_debiasing_test_reweighing,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
temp_stat_par.append(metric_test.mean_difference())
temp_disp_imp.append(metric_test.disparate_impact())
temp_eq_opp_diff.append(metric_test.equal_opportunity_difference())
temp_avg_odds_diff.append(metric_test.average_odds_difference())
temp_theil.append(metric_test.theil_index())
temp_acc.append(acc_test.accuracy())
##################### prejudice remover #####################
prejudice_model_reweighing = PrejudiceRemover(eta=100,
sensitive_attr='sex')
prejudice_model_reweighing.fit(dataset_transf_train)
dataset_prejudice_test_reweighing = prejudice_model_reweighing.predict(
test1)
##################### metrics #####################
metric_test = BinaryLabelDatasetMetric(
dataset_prejudice_test_reweighing,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
acc_test = ClassificationMetric(
test1,
dataset_prejudice_test_reweighing,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
temp_stat_par.append(metric_test.mean_difference())
temp_disp_imp.append(metric_test.disparate_impact())
temp_eq_opp_diff.append(metric_test.equal_opportunity_difference())
temp_avg_odds_diff.append(metric_test.average_odds_difference())
temp_theil.append(metric_test.theil_index())
temp_acc.append(acc_test.accuracy())
##################### normal neural net #####################
sess = tf.Session()
neural_model = AdversarialDebiasing(
privileged_groups=privileged_groups,
unprivileged_groups=unprivileged_groups,
scope_name='debiased_classifier',
debias=False,
sess=sess)
neural_model.fit(dataset_transf_train)
dataset_neural_test = neural_model.predict(test1)
sess.close()
tf.reset_default_graph()
##################### metrics #####################
metric_test = BinaryLabelDatasetMetric(
dataset_neural_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
acc_test = ClassificationMetric(
test1,
dataset_neural_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
temp_stat_par.append(metric_test.mean_difference())
temp_disp_imp.append(metric_test.disparate_impact())
temp_eq_opp_diff.append(metric_test.equal_opportunity_difference())
temp_avg_odds_diff.append(metric_test.average_odds_difference())
temp_theil.append(metric_test.theil_index())
temp_acc.append(acc_test.accuracy())
##################### ensemble #####################
pred_labels_test = []
for i in range(0, len(test1.features)):
arr_test = mode([
dataset_debiasing_test[i], dataset_prejudice_test[i],
dataset_neural_test[i]
])
pred_labels_test.append(arr_test[0][0])
dataset_ensemble_test = test1.copy()
dataset_ensemble_test.labels = np.array(pred_labels_test)
##################### metrics #####################
metric_test = BinaryLabelDatasetMetric(
dataset_ensemble_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
acc_test = ClassificationMetric(
test1,
dataset_ensemble_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
temp_stat_par.append(metric_test.mean_difference())
temp_disp_imp.append(metric_test.disparate_impact())
temp_eq_opp_diff.append(metric_test.equal_opportunity_difference())
temp_avg_odds_diff.append(metric_test.average_odds_difference())
temp_theil.append(metric_test.theil_index())
temp_acc.append(acc_test.accuracy())
######### DUMP SHIT ###########
stat_par_reweigh.append(temp_stat_par)
disp_imp_reweigh.append(temp_disp_imp)
eq_opp_diff_reweigh.append(temp_eq_opp_diff)
avg_odds_diff_reweigh.append(temp_avg_odds_diff)
theil_reweigh.append(temp_theil)
acc_reweigh.append(temp_acc)
without_reweighing = [
stat_par, disp_imp, eq_opp_diff, avg_odds_diff, theil, acc
]
with_reweighing = [
stat_par_reweigh, disp_imp_reweigh, eq_opp_diff_reweigh,
avg_odds_diff_reweigh, theil_reweigh, acc_reweigh
]
for metric in range(len(without_reweighing)):
name = "metric" + str(metric)
sublist = without_reweighing[metric]
with open(name, "wb") as csv_file:
writer = csv.writer(csv_file)
writer.writerows(sublist)
for metric in range(len(with_reweighing)):
name = "metric" + str(metric) + "reweigh"
sublist = with_reweighing[metric]
with open(name, "wb") as csv_file:
writer = csv.writer(csv_file)
writer.writerows(sublist)
dataset_orig_test.features)
dataset_orig_valid.features = min_max_scaler.transform(
dataset_orig_valid.features)
#### PREJUDICE REMOVER ########################################################
pr_predictions_valid = pd.DataFrame()
pr_predictions_test = pd.DataFrame()
all_eta = [1, 5, 15, 30, 50, 70, 100, 150]
for eta in all_eta:
print("Eta: %.2f" % eta)
colname = "eta_" + str(eta)
debiased_model = PrejudiceRemover(eta=eta,
sensitive_attr=protected,
class_attr="TARGET")
debiased_model.fit(dataset_orig_train)
dataset_debiasing_valid = debiased_model.predict(dataset_orig_valid)
dataset_debiasing_test = debiased_model.predict(dataset_orig_test)
scores = dataset_debiasing_valid.scores
pr_predictions_valid[colname] = sum(scores.tolist(), [])
scores = dataset_debiasing_test.scores
pr_predictions_test[colname] = sum(scores.tolist(), [])
pr_predictions_valid.to_csv(output_path + 'taiwan_in_PRpredictions_valid' +
'.csv',
index=None,
def __init__(self, sensitive, class_attr, eta):
self.model = PrejudiceRemover(sensitive_attr=sensitive,
class_attr=class_attr,
eta=eta)