def calc_stat_parity(data, target_variable, protected_variable, unprivileged_input):
df_aif = BinaryLabelDataset(df=data, label_names=[target_variable],
protected_attribute_names=[protected_variable])
privileged_group = []
for v in data[protected_variable].unique()[data[protected_variable].unique() != unprivileged_input]:
privileged_group.append({protected_variable: v})
unprivileged_group = [{protected_variable: unprivileged_input}] #female=0
metric_orig = BinaryLabelDatasetMetric(df_aif, unprivileged_group, privileged_group)
print(metric_orig.statistical_parity_difference().round(3))
if abs(metric_orig.statistical_parity_difference().round(3)) < 0.1:
print('The algorithm can be considered to be not biased')
else:
print('There is a potential bias')
def compute_statistical_parity(data, unpriv_group, priv_group):
if isinstance(data, pd.DataFrame):
transformed_data = BinaryLabelDataset(
df=data,
label_names=["two_year_recid"],
protected_attribute_names=["race"],
favorable_label=0,
unfavorable_label=1)
else:
transformed_data = data
metric_test_data = BinaryLabelDatasetMetric(
transformed_data,
unprivileged_groups=unpriv_group,
privileged_groups=priv_group)
parity_difference = metric_test_data.statistical_parity_difference()
print(
f"Mean difference (statistical parity difference) = {parity_difference}"
)
return parity_difference
def explain(self, request: Dict) -> Dict:
inputs = request["instances"]
predictions = np.array(request["outputs"])
dataframe_predicted = pd.DataFrame(inputs, columns=self.feature_names)
dataframe_predicted[self.label_names[0]] = predictions
dataset_predicted = BinaryLabelDataset(
favorable_label=self.favorable_label,
unfavorable_label=self.unfavorable_label,
df=dataframe_predicted,
label_names=self.label_names,
protected_attribute_names=['age'])
metrics = BinaryLabelDatasetMetric(
dataset_predicted,
unprivileged_groups=self.unprivileged_groups,
privileged_groups=self.privileged_groups)
return {
"predictions": predictions.tolist(),
"metrics": {
"base_rate":
metrics.base_rate(),
"consistency":
metrics.consistency().tolist(),
"disparate_impact":
metrics.disparate_impact(),
"num_instances":
metrics.num_instances(),
"num_negatives":
metrics.num_negatives(),
"num_positives":
metrics.num_positives(),
"statistical_parity_difference":
metrics.statistical_parity_difference(),
}
}
% (train_data.privileged_protected_attributes,
train_data.unprivileged_protected_attributes))
print("Feature names: %s" % train_data.feature_names)
# %% [markdown]
# Now, let's take a look at the test data and compute the following difference:
#
# $$𝑃(𝑌=favorable|𝐷=unprivileged)−𝑃(𝑌=favorable|𝐷=privileged)$$
#
# %%
metric_test_data = BinaryLabelDatasetMetric(test_data,
unprivileged_groups=unpriv_group,
privileged_groups=priv_group)
print("Mean difference (statistical parity difference) = %f" %
metric_test_data.statistical_parity_difference())
def compute_statistical_parity(data, unpriv_group, priv_group):
if isinstance(data, pd.DataFrame):
transformed_data = BinaryLabelDataset(
df=data,
label_names=["two_year_recid"],
protected_attribute_names=["race"],
favorable_label=0,
unfavorable_label=1)
else:
transformed_data = data
metric_test_data = BinaryLabelDatasetMetric(
transformed_data,
def get_bldm_metrics():
metric_BLDM = BinaryLabelDatasetMetric(
dataset, unprivileged_group, privileged_group)
return {"Statistical Parity Difference": metric_BLDM.statistical_parity_difference(), "Disparate Impact": metric_BLDM.disparate_impact()}
