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(),
}
}
def main() -> None:
# read from inventor
filepath = ait_input.get_inventory_path('Data')
# prepare column names as given by german.doc
column_names = [
'status', 'month', 'credit_history', 'purpose', 'credit_amount',
'savings', 'employment', 'investment_as_income_percentage',
'personal_status', 'other_debtors', 'residence_since', 'property',
'age', 'installment_plans', 'housing', 'number_of_credits',
'skill_level', 'people_liable_for', 'telephone', 'foreign_worker',
'credit'
]
# load into a dataframe
df = data_loading(filepath=filepath,
column_names=column_names,
na_values=None)
# prepare for mappings
mappings = {
'label_maps': [{
1.0: 'Good Credit',
2.0: 'Bad Credit'
}],
'protected_attribute_maps': [{
1.0: 'Male',
0.0: 'Female'
}, {
1.0: 'Old',
0.0: 'Young'
}]
}
# prepare for categorical features
categorical_features = [
'status', 'credit_history', 'purpose', 'savings', 'employment',
'other_debtors', 'property', 'installment_plans', 'housing',
'skill_level', 'telephone', 'foreign_worker'
]
# load param
protected_attribute = ait_input.get_method_param_value(
'protected_attribute')
privileged_classes = ait_input.get_method_param_value('privileged_classes')
# input check
ait_input_check(protected_attribute, privileged_classes)
# prepare for structure from dataframe and edit data features setting
dataset = StandardDataset(
df=df,
label_name='credit',
favorable_classes=[1],
protected_attribute_names=[protected_attribute],
privileged_classes=[lambda x: x >= privileged_classes],
instance_weights_name=None,
categorical_features=categorical_features,
features_to_keep=None,
features_to_drop=['personal_status', 'sex'],
na_values=None,
custom_preprocessing=preprocessing,
metadata=mappings)
# set two variables for the privileged (1) and unprivileged (0) values for the age attribute.
privileged_groups = [{protected_attribute: 1}]
unprivileged_groups = [{protected_attribute: 0}]
# compute fairness metric on original training dataset
metric_fairness = BinaryLabelDatasetMetric(
dataset,
unprivileged_groups=unprivileged_groups,
privileged_groups=privileged_groups)
print("Original training dataset: German Credit Data")
print(
"Difference in mean outcomes between unprivileged and privileged groups = %f"
% metric_fairness.mean_difference())
print("unprivileged groups = %f" %
metric_fairness.base_rate(privileged=False))
print("privileged groups = %f" %
metric_fairness.base_rate(privileged=True))
# resource observed_predicted_plot
save_metric_fairness_plot(metric_fairness, protected_attribute)
# measures
measure_mean_difference(metric_fairness.mean_difference())
# ait.log
move_log()
unprivileged_groups=unpriv_group,
privileged_groups=priv_group,
)
tpr_difference = metric_test_data.true_positive_rate_difference()
tpr_priviledged = metric_test_data.true_positive_rate(True)
tpr_unpriviledged = metric_test_data.true_positive_rate(False)
return tpr_difference, tpr_priviledged, tpr_unpriviledged
compute_statistical_parity(test_data.convert_to_dataframe()[0], unpriv_group,
priv_group)
# %% [markdown]
# To be clear, because we're looking at the original label distribution this is the base rate difference between the two groups
# %%
metric_test_data.base_rate(False) # Base rate of the unprivileged group
# %%
metric_test_data.base_rate(True) # Base rate of the privileged group
# %% [markdown]
# To explore the data, it can also help to convert it to a dataframe.
# Note that we get the same numbers as the reported base rates above,
# but because when calculating base rates the favorable label is taken (which is actually 0), it's 1-...
# %% [markdown]
# **Report**
#
# Report basic statistics in your report, such as the size of the training and test set.
#
# Now let's explore the *training* data further.
