def get_classifier_metrics(test_list, prediction_list):
privileged_groups = [{'sex': 1}]
unprivileged_groups = [{'sex': 0}]
acc_list = []
bal_acc_list = []
avg_odds_list = []
recall_diff_list = []
precision_diff_list = []
for test_, pred_ in zip(test_list, prediction_list):
model_metric = ClassificationMetric(
test_,
pred_,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
acc_list.append(model_metric.accuracy().round(3))
bal_acc_list.append(((model_metric.true_positive_rate() +
model_metric.true_negative_rate()) / 2).round(3))
avg_odds_list.append(model_metric.average_odds_difference().round(3))
recall_diff_list.append(
model_metric.equal_opportunity_difference().round(3))
precision_diff_list.append(
(model_metric.precision(privileged=False) -
model_metric.precision(privileged=True)).round(3))
return acc_list, bal_acc_list, avg_odds_list, recall_diff_list, precision_diff_list
def get_metric_reports(true_dataset,classfied_dataset,privileged_groups,unprivileged_groups): mirror_dataset=classfied_dataset.copy(deepcopy=True) mirror_dataset.labels=copy.deepcopy(true_dataset.labels) metric=ClassificationMetric( dataset=mirror_dataset, classified_dataset=classfied_dataset, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) #Measuring unfairness end report=OrderedDict() report['TPR']=metric.true_positive_rate() report['TNR']=metric.true_negative_rate() report['FPR']=metric.false_positive_rate() report['FNR']=metric.false_negative_rate() report['Balanced_Acc']=0.5*(report['TPR']+report['TNR']) report['Acc']=metric.accuracy() report["Statistical parity difference"]=metric.statistical_parity_difference() report["Disparate impact"]=metric.disparate_impact() report["Equal opportunity difference"]=metric.equal_opportunity_difference() report["Average odds difference"]=metric.average_odds_difference() report["Theil index"]=metric.theil_index() report["United Fairness"]=metric.generalized_entropy_index() return report
def compute_metrics(dataset_true,
dataset_pred,
unprivileged_groups,
privileged_groups,
disp=True):
""" Compute the key metrics """
classified_metric_pred = ClassificationMetric(
dataset_true,
dataset_pred,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
metrics = OrderedDict()
metrics["Balanced accuracy"] = 0.5 * (
classified_metric_pred.true_positive_rate() +
classified_metric_pred.true_negative_rate())
metrics[
"Statistical parity difference"] = classified_metric_pred.statistical_parity_difference(
)
metrics["Disparate impact"] = classified_metric_pred.disparate_impact()
metrics[
"Average odds difference"] = classified_metric_pred.average_odds_difference(
)
metrics[
"Equal opportunity difference"] = classified_metric_pred.equal_opportunity_difference(
)
metrics["Theil index"] = classified_metric_pred.theil_index()
if disp:
for k in metrics:
print("%s = %.4f" % (k, metrics[k]))
return metrics
def test(dataset, model, x_test, thresh_arr, unprivileged_groups,
privileged_groups):
bld = BinaryLabelDataset(df=dataset,
label_names=['labels'],
protected_attribute_names=['age'])
if np.isin(k, model_AIF):
y_val_pred_prob = model.predict_proba(bld)
else:
y_val_pred_prob, A_val_pred_prob = model.predict_proba(x_test)
metric_arrs = np.empty([0, 8])
for thresh in thresh_arr:
if np.isin(k, model_AIF):
y_val_pred = (y_val_pred_prob > thresh).astype(np.float64)
else:
y_val_pred = (y_val_pred_prob.numpy() > thresh).astype(np.float64)
metric_arrs = np.append(metric_arrs,
roc_auc_score(y_test, y_val_pred_prob))
if np.isin(k, model_AIF):
metric_arrs = np.append(metric_arrs, 0)
else:
metric_arrs = np.append(metric_arrs,
roc_auc_score(A_test, A_val_pred_prob))
dataset_pred = dataset.copy()
dataset_pred.labels = y_val_pred
bld2 = BinaryLabelDataset(df=dataset_pred,
label_names=['labels'],
protected_attribute_names=['age'])
metric = ClassificationMetric(bld,
bld2,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
metric_arrs = np.append(
metric_arrs,
((metric.true_positive_rate() + metric.true_negative_rate()) / 2))
metric_arrs = np.append(metric_arrs, metric.average_odds_difference())
metric_arrs = np.append(metric_arrs, metric.disparate_impact())
metric_arrs = np.append(metric_arrs,
metric.statistical_parity_difference())
metric_arrs = np.append(metric_arrs,
metric.equal_opportunity_difference())
metric_arrs = np.append(metric_arrs, metric.theil_index())
return metric_arrs
def get_cm_metrics():
df_pred = X.copy()
df_pred[df.columns[-1]] = np.expand_dims(ypred_class, axis=1)
dataset_pred = BinaryLabelDataset(df=df_pred, label_names=[
'action_taken_name'], protected_attribute_names=['applicant_sex_name_Female'])
metric_CM = ClassificationMetric(
dataset, dataset_pred, privileged_groups=privileged_group, unprivileged_groups=unprivileged_group)
return {
"Equal Opportunity Difference": metric_CM.equal_opportunity_difference(),
'Average Odds Difference': metric_CM.average_odds_difference(),
"Accuracy Male": metric_CM.accuracy(privileged=True),
"Accuracy Female": metric_CM.accuracy(privileged=False)
}
def fairness_IBM(y_pred, Ztr, ytr, verbose=0):
from aif360.datasets import BinaryLabelDataset
from aif360.metrics import ClassificationMetric
assert np.array_equal(np.unique(Ztr),
np.array([0, 1])), "Z must contain either 0 or 1"
# if len(ytr.shape) == 1:
# ytr = np.expand_dims(ytr, -1)
Ztr = np.squeeze(Ztr)
if verbose:
print(ytr.shape)
print(Ztr.shape)
unprivileged_groups = [{"zs": [0]}]
privileged_groups = [{"zs": [1]}]
metric_arrs = defaultdict(list)
dict_ = {"y_true": ytr, "zs": Ztr}
df = pd.DataFrame(dict_)
dataset = BinaryLabelDataset(df=df,
label_names=["y_true"],
protected_attribute_names=["zs"],
unprivileged_protected_attributes=[[0]],
privileged_protected_attributes=[[1]])
dataset_pred = dataset.copy()
dataset_pred.labels = y_pred
metric = ClassificationMetric(dataset,
dataset_pred,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
# metric_arrs['bal_acc'].append((metric.true_positive_rate()
# + metric.true_negative_rate()) / 2)
metric_arrs["EA"].append(
metric.accuracy(privileged=False) - metric.accuracy(privileged=True))
# ASSUMING ALL OTHER METRICS RETURN U - P
metric_arrs['EO'].append(metric.average_odds_difference())
# The ideal value of this metric is 1.0
# A value < 1 implies higher benefit for the privileged group
# and a value >1 implies a higher
metric_arrs['DI'].append(metric.disparate_impact() - 1)
metric_arrs['DP'].append(metric.statistical_parity_difference())
metric_arrs['EQ'].append(metric.equal_opportunity_difference())
metric_arrs['TH'].append(metric.between_group_theil_index() * 10)
results = pd.DataFrame(metric_arrs)
return results
def metrics_form(y_val_pred_prob, y_test, A_prob, A_test, bld, dataset):
metric_arrs = np.empty([0, 8])
if np.isin(k, model_AIF):
y_val_pred = (y_val_pred_prob > thresh).astype(np.float64)
else:
y_val_pred = (y_val_pred_prob > thresh).astype(np.float64)
A_pred = (A_prob > thresh).astype(np.float64)
metric_arrs = np.append(metric_arrs,
roc_auc_score(y_test, y_val_pred_prob))
print("y {}".format(roc_auc_score(y_test, y_val_pred_prob)))
metric_arrs = np.append(metric_arrs,
accuracy_score(y_test, y_val_pred))
if np.isin(k, model_AIF):
metric_arrs = np.append(metric_arrs, 0)
else:
metric_arrs = np.append(metric_arrs, roc_auc_score(A_test, A_prob))
print("A {}".format(roc_auc_score(A_test, A_prob)))
dataset_pred = dataset.copy()
dataset_pred.labels = y_val_pred
bld2 = BinaryLabelDataset(df=dataset_pred,
label_names=['labels'],
protected_attribute_names=protected)
metric = ClassificationMetric(bld,
bld2,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
metric_arrs = np.append(
metric_arrs,
((metric.true_positive_rate() + metric.true_negative_rate()) / 2))
metric_arrs = np.append(metric_arrs,
np.abs(metric.average_odds_difference()))
metric_arrs = np.append(metric_arrs, metric.disparate_impact())
metric_arrs = np.append(metric_arrs,
np.abs(metric.statistical_parity_difference()))
metric_arrs = np.append(metric_arrs,
np.abs(metric.equal_opportunity_difference()))
return metric_arrs
def compute_aif_metrics(dataset_true, dataset_pred, unprivileged_groups, privileged_groups,\
ret_eval_dict=True):
metrics_cls = ClassificationMetric(dataset_true, dataset_pred,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
metrics_dict = {}
metrics_dict["BA"] = 0.5*(metrics_cls.true_positive_rate()+
metrics_cls.true_negative_rate())
metrics_dict["SPD"] = metrics_cls.statistical_parity_difference()
metrics_dict["DI"] = metrics_cls.disparate_impact()
metrics_dict["AOD"] = metrics_cls.average_odds_difference()
metrics_dict["EOD"] = metrics_cls.equal_opportunity_difference()
metrics_dict["DFBA"] = metrics_cls.differential_fairness_bias_amplification()
metrics_dict["TI"] = metrics_cls.theil_index()
if ret_eval_dict:
return metrics_dict, metrics_cls
else:
return metrics_cls
"#### Plain model - without debiasing - classification metrics on test set"
)
# print("Test set: Classification accuracy = %f" % classified_metric_nodebiasing_test.accuracy())
# print("Test set: Balanced classification accuracy = %f" % bal_acc_nodebiasing_test)
# print("Test set: Statistical parity difference = %f" % classified_metric_nodebiasing_test.statistical_parity_difference())
# print("Test set: Disparate impact = %f" % classified_metric_nodebiasing_test.disparate_impact())
# print("Test set: Equal opportunity difference = %f" % classified_metric_nodebiasing_test.equal_opportunity_difference())
# print("Test set: Average odds difference = %f" % classified_metric_nodebiasing_test.average_odds_difference())
# print("Test set: Theil index = %f" % classified_metric_nodebiasing_test.theil_index())
# print("Test set: False negative rate difference = %f" % classified_metric_nodebiasing_test.false_negative_rate_difference())
metrics = {
"Classification accuracy":
classified_metric_nodebiasing_test.accuracy(),
"Balanced classification accuracy":
bal_acc_nodebiasing_test,
"Statistical parity difference":
classified_metric_nodebiasing_test.statistical_parity_difference(),
"Disparate impact":
classified_metric_nodebiasing_test.disparate_impact(),
"Equal opportunity difference":
classified_metric_nodebiasing_test.equal_opportunity_difference(),
"Average odds difference":
classified_metric_nodebiasing_test.average_odds_difference(),
"Theil index":
classified_metric_nodebiasing_test.theil_index(),
"False negative rate difference":
classified_metric_nodebiasing_test.false_negative_rate_difference()
}
print("metrics: ", metrics)
equal_opp_diff = []
for c in C:
predictions, _ = train_and_predict(X_train, y_train, X_test, c, norm_type)
ds_te_pred = ds_te.copy()
ds_te_pred.labels = predictions
metric_te = ClassificationMetric(ds_te,
ds_te_pred,
unprivileged_groups=unpriv,
privileged_groups=priv)
BACC = 0.5*(metric_te.true_positive_rate()\
+metric_te.true_negative_rate())
metric_1 = metric_te.statistical_parity_difference()
metric_2 = metric_te.average_odds_difference()
metric_3 = metric_te.equal_opportunity_difference()
accuracy.append(BACC)
mean_diff.append(metric_1)
average_odds_diff.append(metric_2)
equal_opp_diff.append(metric_3)
# save plots
plot_results(C, norm_type, accuracy, mean_diff, average_odds_diff, \
equal_opp_diff, name+'_all_metrics_'+norm_type)
def results_table(C, accuracy, mean_diff, avg_odds_diff, equal_opp_diff):
results = pd.DataFrame()
results['c'] = C
ClassificationMetric(
dataset_ground_truth,
dataset_classifier,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
TPR = classificaltion_metric.true_positive_rate()
TNR = classificaltion_metric.true_negative_rate()
bal_acc_nodebiasing_test = 0.5 * (TPR + TNR)
metrics = {
"classification_accuracy":
classificaltion_metric.accuracy(),
"balanced_classification_accuracy":
bal_acc_nodebiasing_test,
"statistical_parity_difference":
classificaltion_metric.statistical_parity_difference(),
"disparate_impact":
classificaltion_metric.disparate_impact(),
"equal_opportunity_difference":
classificaltion_metric.equal_opportunity_difference(),
"average_odds_difference":
classificaltion_metric.average_odds_difference(),
"theil_index":
classificaltion_metric.theil_index(),
"false_negative_rate_difference":
classificaltion_metric.false_negative_rate_difference()
}
sys.stdout.write(json.dumps(metrics))
#%%
# =============================================================================
# FAIRNESS METRICS (FOR REAL THIS TIME)
# =============================================================================
# TODO: (1) Redo the previous cell for gender bias and recompute the corresponding
# fairness metrics
# (2)collect these values in a table
# (3) think about a way to visualize these values
# Statistical Parity difference (SPD)
spd_pre_race = fairness_metrics.statistical_parity_difference()
# Disparate Impact Ratio
dir_pre_race = fairness_metrics.disparate_impact()
# Average Odds Difference and Average absolute odds difference
aod_pre_race = fairness_metrics.average_odds_difference()
aaod_pre_race = fairness_metrics.average_abs_odds_difference()
# Equal Opportunity Difference aka true positive rate difference
eod_pre_race = fairness_metrics.equal_opportunity_difference()
# Generealized entropy index with various alpha's
fairness_metrics.between_all_groups_generalized_entropy_index(alpha=2)
ClassificationMetric(dataset=bld_true,
classified_dataset=bld_pred,
unprivileged_groups=None,
privileged_groups=None).false_positive_rate()
df_fm.head()
def fairness_check(s3_url, bucket_name, s3_username, s3_password, training_id):
cos = boto3.resource("s3",
endpoint_url=s3_url,
aws_access_key_id=s3_username,
aws_secret_access_key=s3_password)
y_test_out = 'y_test.out'
p_test_out = 'p_test.out'
y_pred_out = 'y_pred.out'
get_s3_item(cos, bucket_name, training_id + '/' + y_test_out, y_test_out)
get_s3_item(cos, bucket_name, training_id + '/' + p_test_out, p_test_out)
get_s3_item(cos, bucket_name, training_id + '/' + y_pred_out, y_pred_out)
"""Need to generalize the protected features"""
unprivileged_groups = [{'race': 4.0}]
privileged_groups = [{'race': 0.0}]
favorable_label = 0.0
unfavorable_label = 1.0
"""Load the necessary labels and protected features for fairness check"""
y_test = np.loadtxt(y_test_out)
p_test = np.loadtxt(p_test_out)
y_pred = np.loadtxt(y_pred_out)
"""Calculate the fairness metrics"""
original_test_dataset = dataset_wrapper(outcome=y_test, protected=p_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups,
favorable_label=favorable_label,
unfavorable_label=unfavorable_label)
plain_predictions_test_dataset = dataset_wrapper(outcome=y_pred, protected=p_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups,
favorable_label=favorable_label,
unfavorable_label=unfavorable_label)
classified_metric_nodebiasing_test = ClassificationMetric(original_test_dataset,
plain_predictions_test_dataset,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
TPR = classified_metric_nodebiasing_test.true_positive_rate()
TNR = classified_metric_nodebiasing_test.true_negative_rate()
bal_acc_nodebiasing_test = 0.5*(TPR+TNR)
print("#### Plain model - without debiasing - classification metrics on test set")
metrics = {
"Classification accuracy": classified_metric_nodebiasing_test.accuracy(),
"Balanced classification accuracy": bal_acc_nodebiasing_test,
"Statistical parity difference": classified_metric_nodebiasing_test.statistical_parity_difference(),
"Disparate impact": classified_metric_nodebiasing_test.disparate_impact(),
"Equal opportunity difference": classified_metric_nodebiasing_test.equal_opportunity_difference(),
"Average odds difference": classified_metric_nodebiasing_test.average_odds_difference(),
"Theil index": classified_metric_nodebiasing_test.theil_index(),
"False negative rate difference": classified_metric_nodebiasing_test.false_negative_rate_difference()
}
print("metrics: ", metrics)
return metrics
def comb_algorithm(l, m, n, dataset_original1, privileged_groups1,
unprivileged_groups1, optim_options1):
dataset_original2 = copy.deepcopy(dataset_original1)
privileged_groups2 = copy.deepcopy(privileged_groups1)
unprivileged_groups2 = copy.deepcopy(unprivileged_groups1)
optim_options2 = copy.deepcopy(optim_options1)
print(l, m, n)
dataset_original_train, dataset_original_vt = dataset_original2.split(
[0.7], shuffle=True)
dataset_original_valid, dataset_original_test = dataset_original_vt.split(
[0.5], shuffle=True)
dataset_original_test.labels = dataset_original_test.labels
print('=======================')
#print(dataset_original_test.labels)
dataset_orig_train = copy.deepcopy(dataset_original_train)
dataset_orig_valid = copy.deepcopy(dataset_original_valid)
dataset_orig_test = copy.deepcopy(dataset_original_test)
if l == 0:
dataset_transfer_train = copy.deepcopy(dataset_original_train)
dataset_transfer_valid = copy.deepcopy(dataset_original_valid)
dataset_transfer_test = copy.deepcopy(dataset_original_test)
#dataset_transf_train, dataset_transf_valid, dataset_transf_test = dataset_orig_train, dataset_orig_valid, dataset_orig_test
else:
pre_used = preAlgorithm[l - 1]
dataset_transfer_train, dataset_transfer_valid, dataset_transfer_test = Pre(
pre_used, dataset_orig_train, dataset_orig_valid,
dataset_orig_test, privileged_groups2, unprivileged_groups2,
optim_options2)
dataset_transf_train = copy.deepcopy(dataset_transfer_train)
dataset_transf_valid = copy.deepcopy(dataset_transfer_valid)
dataset_transf_test = copy.deepcopy(dataset_transfer_test)
if m == 0:
dataset_transfer_valid_pred, dataset_transfer_test_pred = plain_model(
dataset_transf_train, dataset_transf_valid, dataset_transf_test,
privileged_groups2, unprivileged_groups2)
else:
in_used = inAlgorithm[m - 1]
if in_used == "adversarial_debiasing":
dataset_transfer_valid_pred, dataset_transfer_test_pred = adversarial_debiasing(
dataset_transf_train, dataset_transf_valid,
dataset_transf_test, privileged_groups2, unprivileged_groups2)
elif in_used == "art_classifier":
dataset_transfer_valid_pred, dataset_transfer_test_pred = art_classifier(
dataset_transf_train, dataset_transf_valid,
dataset_transf_test, privileged_groups2, unprivileged_groups2)
elif in_used == "prejudice_remover":
for key, value in privileged_groups2[0].items():
sens_attr = key
dataset_transfer_valid_pred, dataset_transfer_test_pred = prejudice_remover(
dataset_transf_train, dataset_transf_valid,
dataset_transf_test, privileged_groups2, unprivileged_groups2,
sens_attr)
dataset_transf_valid_pred = copy.deepcopy(dataset_transfer_valid_pred)
dataset_transf_test_pred = copy.deepcopy(dataset_transfer_test_pred)
if n == 0:
dataset_transf_test_pred_transf = copy.deepcopy(
dataset_transfer_test_pred)
else:
post_used = postAlgorithm[n - 1]
if post_used == "calibrated_eqodds":
cpp = CalibratedEqOddsPostprocessing(
privileged_groups=privileged_groups2,
unprivileged_groups=unprivileged_groups2,
cost_constraint=cost_constraint)
cpp = cpp.fit(dataset_transfer_valid, dataset_transf_valid_pred)
dataset_transf_test_pred_transf = cpp.predict(
dataset_transf_test_pred)
elif post_used == "eqodds":
EO = EqOddsPostprocessing(unprivileged_groups=unprivileged_groups2,
privileged_groups=privileged_groups2)
EO = EO.fit(dataset_transfer_valid, dataset_transf_valid_pred)
dataset_transf_test_pred_transf = EO.predict(
dataset_transf_test_pred)
elif post_used == "reject_option":
#dataset_transf_test_pred_transf = reject_option(dataset_transf_valid, dataset_transf_valid_pred, dataset_transf_test, dataset_transf_test_pred, privileged_groups2, unprivileged_groups2)
ROC = RejectOptionClassification(
unprivileged_groups=unprivileged_groups2,
privileged_groups=privileged_groups2)
ROC = ROC.fit(dataset_transfer_valid, dataset_transf_valid_pred)
dataset_transf_test_pred_transf = ROC.predict(
dataset_transf_test_pred)
#print('=======================')
org_labels = dataset_orig_test.labels
print(dataset_orig_test.labels)
#print(dataset_transf_test.labels)
#print('=======================')
pred_labels = dataset_transf_test_pred.labels
print(dataset_transf_test_pred.labels)
true_pred = org_labels == pred_labels
print("acc after in: ", float(np.sum(true_pred)) / pred_labels.shape[1])
#print('=======================')
#print(dataset_transf_test_pred_transf.labels)
#print(dataset_transf_test_pred_transf.labels.shape)
metric = ClassificationMetric(dataset_transfer_test,
dataset_transf_test_pred_transf,
unprivileged_groups=unprivileged_groups2,
privileged_groups=privileged_groups2)
metrics = OrderedDict()
metrics["Classification accuracy"] = metric.accuracy()
TPR = metric.true_positive_rate()
TNR = metric.true_negative_rate()
bal_acc_nodebiasing_test = 0.5 * (TPR + TNR)
metrics["Balanced classification accuracy"] = bal_acc_nodebiasing_test
metrics[
"Statistical parity difference"] = metric.statistical_parity_difference(
)
metrics["Disparate impact"] = metric.disparate_impact()
metrics[
"Equal opportunity difference"] = metric.equal_opportunity_difference(
)
metrics["Average odds difference"] = metric.average_odds_difference()
metrics["Theil index"] = metric.theil_index()
metrics["United Fairness"] = metric.generalized_entropy_index()
feature = []
feature_str = "["
for m in metrics:
data = round(metrics[m], 4)
feature.append(data)
feature_str = feature_str + str(data) + " "
feature_str = feature_str + "]"
return feature, feature_str
def compute_metrics(model, X_test, y_test, X_train, y_train, dataset_test,
dataset_name, model_name, unprivileged_groups,
privileged_groups, position):
"""
Calculate and return: model accuracy and fairness metrics
Parameters
----------
model: scikit-learn classifier
X_test: numpy 2d array
y_test: numpy 1d array
X_train: numpy 2d array
y_train: numpy 1d array
dataset_test: aif360.datasets.BinaryLabelDataset
dataset_name: string
Dataset name used in the analysis
model_name: string
unprivileged_groups: list<dict>
Dictionary where the key is the name of the sensitive column in the
dataset, and the value is the value of the unprivileged group in the
dataset
privileged_groups: list<dict>
Dictionary where the key is the name of the sensitive column in the
dataset, and the value is the value of the privileged group in the
dataset
position: int
Column position of the sensitive group in the dataset
"""
y_pred_test = model.predict(X_test)
acc_test = accuracy_score(y_true=y_test, y_pred=y_pred_test)
print("Test accuracy: ", acc_test)
y_pred_train = model.predict(X_train)
acc_train = accuracy_score(y_true=y_train, y_pred=y_pred_train)
print("Train accuracy: ", acc_train)
dataset_pred = dataset_test.copy()
dataset_pred.labels = y_pred_test
bin_metric = BinaryLabelDatasetMetric(
dataset_pred,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
disparate_impact_bin = bin_metric.disparate_impact()
print('Disparate impact: ', disparate_impact_bin)
mean_difference = bin_metric.mean_difference()
print('Mean difference: ', mean_difference)
classif_metric = ClassificationMetric(
dataset_test,
dataset_pred,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
classif_disparete_impact = classif_metric.disparate_impact()
avg_odds = classif_metric.average_odds_difference()
print('Average odds difference:', avg_odds)
equal_opport = classif_metric.equal_opportunity_difference()
print('Equality of opportunity:', equal_opport)
false_discovery_rate = classif_metric.false_discovery_rate_difference()
print('False discovery rate difference:', false_discovery_rate)
entropy_index = classif_metric.generalized_entropy_index()
print('Generalized entropy index:', entropy_index)
cons_comp = consitency_mod(bin_metric, position, n_neighbors=5)
print('Consistency: ', cons_comp)
result = (dataset_name, model_name, acc_test, disparate_impact_bin,
mean_difference, classif_disparete_impact, avg_odds,
equal_opport, false_discovery_rate, entropy_index, cons_comp)
return result
def fairness_check(label_dir, model_dir):
"""Need to generalize the protected features"""
# races_to_consider = [0,4]
unprivileged_groups = [{'race': 4.0}]
privileged_groups = [{'race': 0.0}]
favorable_label = 0.0
unfavorable_label = 1.0
"""Load the necessary labels and protected features for fairness check"""
# y_train = np.loadtxt(label_dir + '/y_train.out')
# p_train = np.loadtxt(label_dir + '/p_train.out')
y_test = np.loadtxt(label_dir + '/y_test.out')
p_test = np.loadtxt(label_dir + '/p_test.out')
y_pred = np.loadtxt(label_dir + '/y_pred.out')
"""Calculate the fairness metrics"""
# original_traning_dataset = dataset_wrapper(outcome=y_train, protected=p_train,
# unprivileged_groups=unprivileged_groups,
# privileged_groups=privileged_groups,
# favorable_label=favorable_label,
# unfavorable_label=unfavorable_label)
original_test_dataset = dataset_wrapper(outcome=y_test, protected=p_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups,
favorable_label=favorable_label,
unfavorable_label=unfavorable_label)
plain_predictions_test_dataset = dataset_wrapper(outcome=y_pred, protected=p_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups,
favorable_label=favorable_label,
unfavorable_label=unfavorable_label)
classified_metric_nodebiasing_test = ClassificationMetric(original_test_dataset,
plain_predictions_test_dataset,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
TPR = classified_metric_nodebiasing_test.true_positive_rate()
TNR = classified_metric_nodebiasing_test.true_negative_rate()
bal_acc_nodebiasing_test = 0.5*(TPR+TNR)
print("#### Plain model - without debiasing - classification metrics on test set")
# print("Test set: Classification accuracy = %f" % classified_metric_nodebiasing_test.accuracy())
# print("Test set: Balanced classification accuracy = %f" % bal_acc_nodebiasing_test)
# print("Test set: Statistical parity difference = %f" % classified_metric_nodebiasing_test.statistical_parity_difference())
# print("Test set: Disparate impact = %f" % classified_metric_nodebiasing_test.disparate_impact())
# print("Test set: Equal opportunity difference = %f" % classified_metric_nodebiasing_test.equal_opportunity_difference())
# print("Test set: Average odds difference = %f" % classified_metric_nodebiasing_test.average_odds_difference())
# print("Test set: Theil index = %f" % classified_metric_nodebiasing_test.theil_index())
# print("Test set: False negative rate difference = %f" % classified_metric_nodebiasing_test.false_negative_rate_difference())
metrics = {
"Classification accuracy": classified_metric_nodebiasing_test.accuracy(),
"Balanced classification accuracy": bal_acc_nodebiasing_test,
"Statistical parity difference": classified_metric_nodebiasing_test.statistical_parity_difference(),
"Disparate impact": classified_metric_nodebiasing_test.disparate_impact(),
"Equal opportunity difference": classified_metric_nodebiasing_test.equal_opportunity_difference(),
"Average odds difference": classified_metric_nodebiasing_test.average_odds_difference(),
"Theil index": classified_metric_nodebiasing_test.theil_index(),
"False negative rate difference": classified_metric_nodebiasing_test.false_negative_rate_difference()
}
return {"metrics": metrics}
def fit(self, dataset_true, dataset_pred):
"""Estimates the optimal classification threshold and margin for reject
option classification that optimizes the metric provided.
Note:
The `fit` function is a no-op for this algorithm.
Args:
dataset_true (BinaryLabelDataset): Dataset containing the true
`labels`.
dataset_pred (BinaryLabelDataset): Dataset containing the predicted
`scores`.
Returns:
RejectOptionClassification: Returns self.
"""
fair_metric_arr = np.zeros(self.num_class_thresh*self.num_ROC_margin)
balanced_acc_arr = np.zeros_like(fair_metric_arr)
ROC_margin_arr = np.zeros_like(fair_metric_arr)
class_thresh_arr = np.zeros_like(fair_metric_arr)
cnt = 0
# Iterate through class thresholds
for class_thresh in np.linspace(self.low_class_thresh,
self.high_class_thresh,
self.num_class_thresh):
self.classification_threshold = class_thresh
if class_thresh <= 0.5:
low_ROC_margin = 0.0
high_ROC_margin = class_thresh
else:
low_ROC_margin = 0.0
high_ROC_margin = (1.0-class_thresh)
# Iterate through ROC margins
for ROC_margin in np.linspace(
low_ROC_margin,
high_ROC_margin,
self.num_ROC_margin):
self.ROC_margin = ROC_margin
# Predict using the current threshold and margin
dataset_transf_pred = self.predict(dataset_pred)
dataset_transf_metric_pred = BinaryLabelDatasetMetric(
dataset_transf_pred,
unprivileged_groups=self.unprivileged_groups,
privileged_groups=self.privileged_groups)
classified_transf_metric = ClassificationMetric(
dataset_true,
dataset_transf_pred,
unprivileged_groups=self.unprivileged_groups,
privileged_groups=self.privileged_groups)
ROC_margin_arr[cnt] = self.ROC_margin
class_thresh_arr[cnt] = self.classification_threshold
# Balanced accuracy and fairness metric computations
balanced_acc_arr[cnt] = 0.5*(classified_transf_metric.true_positive_rate()\
+classified_transf_metric.true_negative_rate())
if self.metric_name == "Statistical parity difference":
fair_metric_arr[cnt] = dataset_transf_metric_pred.mean_difference()
elif self.metric_name == "Average odds difference":
fair_metric_arr[cnt] = classified_transf_metric.average_odds_difference()
elif self.metric_name == "Equal opportunity difference":
fair_metric_arr[cnt] = classified_transf_metric.equal_opportunity_difference()
cnt += 1
rel_inds = np.logical_and(fair_metric_arr >= self.metric_lb,
fair_metric_arr <= self.metric_ub)
if any(rel_inds):
best_ind = np.where(balanced_acc_arr[rel_inds]
== np.max(balanced_acc_arr[rel_inds]))[0][0]
else:
warn("Unable to satisy fairness constraints")
rel_inds = np.ones(len(fair_metric_arr), dtype=bool)
best_ind = np.where(fair_metric_arr[rel_inds]
== np.min(fair_metric_arr[rel_inds]))[0][0]
self.ROC_margin = ROC_margin_arr[rel_inds][best_ind]
self.classification_threshold = class_thresh_arr[rel_inds][best_ind]
return self
def compute_metrics(model, X_test, y_test, X_train, y_train, dataset_test,
unprivileged_groups, privileged_groups, protect_attribute,
print_result):
"""
Calculate and return: model accuracy and fairness metrics
Parameters
----------
model: scikit-learn classifier
X_test: numpy 2d array
y_test: numpy 1d array
X_train: numpy 2d array
y_train: numpy 1d array
dataset_test: aif360.datasets.BinaryLabelDataset
unprivileged_groups: list<dict>
Dictionary where the key is the name of the sensitive column in the
dataset, and the value is the value of the unprivileged group in the
dataset
privileged_groups: list<dict>
Dictionary where the key is the name of the sensitive column in the
dataset, and the value is the value of the privileged group in the
dataset
protect_attribute
print_result
"""
result = {}
y_pred_test = model.predict(X_test)
result['acc_test'] = accuracy_score(y_true=y_test, y_pred=y_pred_test)
y_pred_train = model.predict(X_train)
result['acc_train'] = accuracy_score(y_true=y_train, y_pred=y_pred_train)
dataset_pred = dataset_test.copy()
dataset_pred.labels = y_pred_test
bin_metric = BinaryLabelDatasetMetric(dataset_pred,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
result['disp_impact'] = bin_metric.disparate_impact()
result['stat_parity'] = bin_metric.mean_difference()
classif_metric = ClassificationMetric(dataset_test, dataset_pred,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
result['avg_odds'] = classif_metric.average_odds_difference()
result['equal_opport'] = classif_metric.equal_opportunity_difference()
result['false_discovery_rate'] = classif_metric.false_discovery_rate_difference()
result['entropy_index'] = classif_metric.generalized_entropy_index()
result['acc_test_clf'] = classif_metric.accuracy(privileged=None)
result['acc_test_priv'] = classif_metric.accuracy(privileged=True)
result['acc_test_unpriv'] = classif_metric.accuracy(privileged=False)
result['consistency'] = consitency(X_test, y_pred_test, protect_attribute, n_neighbors=5)
result['counterfactual'] = counterfactual(X_test, model, protect_attribute)
if print_result:
print("Train accuracy: ", result['acc_train'])
print("Test accuracy: ", result['acc_test'])
print("Test accuracy clf: ", result['acc_test_clf'])
print("Test accuracy priv.: ", result['acc_test_priv'])
print("Test accuracy unpriv.: ", result['acc_test_unpriv'])
print('Disparate impact: ', result['disp_impact'])
print('Mean difference: ', result['stat_parity'])
print('Average odds difference:', result['avg_odds'])
print('Equality of opportunity:', result['equal_opport'])
print('False discovery rate difference:', result['false_discovery_rate'])
print('Generalized entropy index:', result['entropy_index'])
print('Consistency: ', result['consistency'])
print('Counterfactual fairness: ', result['counterfactual'])
return result
def fairness_check(object_storage_url,
object_storage_username,
object_storage_password,
data_bucket_name,
result_bucket_name,
model_id,
feature_testset_path='processed_data/X_test.npy',
label_testset_path='processed_data/y_test.npy',
protected_label_testset_path='processed_data/p_test.npy',
model_class_file='model.py',
model_class_name='model',
favorable_label=0.0,
unfavorable_label=1.0,
privileged_groups=[{
'race': 0.0
}],
unprivileged_groups=[{
'race': 4.0
}]):
url = re.compile(r"https?://")
cos = Minio(url.sub('', object_storage_url),
access_key=object_storage_username,
secret_key=object_storage_password,
secure=False) # Local Minio server won't have HTTPS
dataset_filenamex = "X_test.npy"
dataset_filenamey = "y_test.npy"
dataset_filenamep = "p_test.npy"
weights_filename = "model.pt"
model_files = model_id + '/_submitted_code/model.zip'
cos.fget_object(data_bucket_name, feature_testset_path, dataset_filenamex)
cos.fget_object(data_bucket_name, label_testset_path, dataset_filenamey)
cos.fget_object(data_bucket_name, protected_label_testset_path,
dataset_filenamep)
cos.fget_object(result_bucket_name, model_id + '/' + weights_filename,
weights_filename)
cos.fget_object(result_bucket_name, model_files, 'model.zip')
# Load PyTorch model definition from the source code.
zip_ref = zipfile.ZipFile('model.zip', 'r')
zip_ref.extractall('model_files')
zip_ref.close()
modulename = 'model_files.' + model_class_file.split('.')[0].replace(
'-', '_')
'''
We required users to define where the model class is located or follow
some naming convention we have provided.
'''
model_class = getattr(importlib.import_module(modulename),
model_class_name)
# load & compile model
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
model = model_class().to(device)
model.load_state_dict(torch.load(weights_filename, map_location=device))
"""Load the necessary labels and protected features for fairness check"""
x_test = np.load(dataset_filenamex)
y_test = np.load(dataset_filenamey)
p_test = np.load(dataset_filenamep)
_, y_pred = evaluate(model, x_test, y_test)
"""Calculate the fairness metrics"""
original_test_dataset = dataset_wrapper(
outcome=y_test,
protected=p_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups,
favorable_label=favorable_label,
unfavorable_label=unfavorable_label)
plain_predictions_test_dataset = dataset_wrapper(
outcome=y_pred,
protected=p_test,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups,
favorable_label=favorable_label,
unfavorable_label=unfavorable_label)
classified_metric_nodebiasing_test = ClassificationMetric(
original_test_dataset,
plain_predictions_test_dataset,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
TPR = classified_metric_nodebiasing_test.true_positive_rate()
TNR = classified_metric_nodebiasing_test.true_negative_rate()
bal_acc_nodebiasing_test = 0.5 * (TPR + TNR)
print(
"#### Plain model - without debiasing - classification metrics on test set"
)
metrics = {
"Classification accuracy":
classified_metric_nodebiasing_test.accuracy(),
"Balanced classification accuracy":
bal_acc_nodebiasing_test,
"Statistical parity difference":
classified_metric_nodebiasing_test.statistical_parity_difference(),
"Disparate impact":
classified_metric_nodebiasing_test.disparate_impact(),
"Equal opportunity difference":
classified_metric_nodebiasing_test.equal_opportunity_difference(),
"Average odds difference":
classified_metric_nodebiasing_test.average_odds_difference(),
"Theil index":
classified_metric_nodebiasing_test.theil_index(),
"False negative rate difference":
classified_metric_nodebiasing_test.false_negative_rate_difference()
}
print("metrics: ", metrics)
return metrics
def comb_algorithm(l, m, n, dataset_original1, privileged_groups1,
unprivileged_groups1, optim_options1):
dataset_original2 = copy.deepcopy(dataset_original1)
privileged_groups2 = copy.deepcopy(privileged_groups1)
unprivileged_groups2 = copy.deepcopy(unprivileged_groups1)
optim_options2 = copy.deepcopy(optim_options1)
print(l, m, n)
dataset_orig_train, dataset_orig_vt = dataset_original2.split([0.7],
shuffle=True)
dataset_orig_valid, dataset_orig_test = dataset_orig_vt.split([0.5],
shuffle=True)
if l == 0:
dataset_transf_train, dataset_transf_valid, dataset_transf_test = dataset_orig_train, dataset_orig_valid, dataset_orig_test
else:
pre_used = preAlgorithm[l - 1]
dataset_transf_train, dataset_transf_valid, dataset_transf_test = Pre(
pre_used, dataset_orig_train, dataset_orig_valid,
dataset_orig_test, privileged_groups2, unprivileged_groups2,
optim_options2)
#assert (l,m,n)!=(2,0,0)
#assert not np.all(dataset_transf_train.labels.flatten()==1.0)
if m == 0:
dataset_transf_valid_pred, dataset_transf_test_pred = train(
dataset_transf_train, dataset_transf_valid, dataset_transf_test,
privileged_groups2, unprivileged_groups2)
else:
in_used = inAlgorithm[m - 1]
if in_used == "adversarial_debiasing":
dataset_transf_valid_pred, dataset_transf_test_pred = adversarial_debiasing(
dataset_transf_train, dataset_transf_valid,
dataset_transf_test, privileged_groups2, unprivileged_groups2)
elif in_used == "art_classifier":
dataset_transf_valid_pred, dataset_transf_test_pred = art_classifier(
dataset_transf_train, dataset_transf_valid,
dataset_transf_test, privileged_groups2, unprivileged_groups2)
elif in_used == "prejudice_remover":
for key, value in privileged_groups2[0].items():
sens_attr = key
dataset_transf_valid_pred, dataset_transf_test_pred = prejudice_remover(
dataset_transf_train, dataset_transf_valid,
dataset_transf_test, privileged_groups2, unprivileged_groups2,
sens_attr)
if n == 0:
dataset_transf_test_pred_transf = dataset_transf_test_pred
else:
post_used = postAlgorithm[n - 1]
if post_used == "calibrated_eqodds":
cpp = CalibratedEqOddsPostprocessing(
privileged_groups=privileged_groups2,
unprivileged_groups=unprivileged_groups2,
cost_constraint=cost_constraint,
seed=1)
cpp = cpp.fit(dataset_transf_valid, dataset_transf_valid_pred)
dataset_transf_test_pred_transf = cpp.predict(
dataset_transf_test_pred)
elif post_used == "eqodds":
EO = EqOddsPostprocessing(unprivileged_groups=unprivileged_groups2,
privileged_groups=privileged_groups2,
seed=1)
EO = EO.fit(dataset_transf_valid, dataset_transf_valid_pred)
dataset_transf_test_pred_transf = EO.predict(
dataset_transf_test_pred)
elif post_used == "reject_option":
ROC = RejectOptionClassification(
unprivileged_groups=unprivileged_groups2,
privileged_groups=privileged_groups2,
low_class_thresh=0.01,
high_class_thresh=0.99,
num_class_thresh=100,
num_ROC_margin=50,
metric_name=allowed_metrics[0],
metric_ub=metric_ub,
metric_lb=metric_lb)
ROC = ROC.fit(dataset_transf_valid, dataset_transf_valid_pred)
dataset_transf_test_pred_transf = ROC.predict(
dataset_transf_test_pred)
metric = ClassificationMetric(dataset_transf_test,
dataset_transf_test_pred_transf,
unprivileged_groups=unprivileged_groups2,
privileged_groups=privileged_groups2)
metrics = OrderedDict()
metrics["Classification accuracy"] = metric.accuracy()
TPR = metric.true_positive_rate()
TNR = metric.true_negative_rate()
bal_acc_nodebiasing_test = 0.5 * (TPR + TNR)
metrics["Balanced classification accuracy"] = bal_acc_nodebiasing_test
metrics[
"Statistical parity difference"] = metric.statistical_parity_difference(
)
metrics["Disparate impact"] = metric.disparate_impact()
metrics[
"Equal opportunity difference"] = metric.equal_opportunity_difference(
)
metrics["Average odds difference"] = metric.average_odds_difference()
metrics["Theil index"] = metric.theil_index()
metrics["United Fairness"] = metric.generalized_entropy_index()
# print(metrics)
feature = "["
for m in metrics:
feature = feature + " " + str(round(metrics[m], 4))
feature = feature + "]"
return feature
