def fairness_dashboard(sensitive_feature_df, yelp_test_dataset):
from fairlearn.widget import FairlearnDashboard
#TODO split out test dataset to features and labels
X_test, y_test = yelp_test_dataset, yelp_test_dataset
FairlearnDashboard(sensitive_features=sensitive_feature_df,
sensitive_feature_names=['gender'],
y_true=y_test,
y_pred={"lr_model": predict(text, model, vocab, ngrams=2)})
def show_comparison(model, X_test, y_test, A_test, protected_features,
prostprocess_preds):
"""
Returns Dashboard to show comparison of models based on the trade off of the disparity and accuracy
"""
FairlearnDashboard(sensitive_features=A_test,
sensitive_feature_names=protected_features,
y_true=Y_test,
y_pred={
"Unmitigated": model.predict(X_test),
"ThresholdOptimizer": postprocess_preds
})
return dashboard
def show_comparison(model, X_test, Y_test, A_test, protected_features,
non_dominated):
"""
Returns Dashboard to show comparison of models based on the trade off of the disparity and accuracy
"""
dashboard_predicted = {"unmitigated": model.predict(X_test)}
for i in range(len(non_dominated)):
key = "dominant_model_{0}".format(i)
value = non_dominated[i].predict(X_test)
dashboard_predicted[key] = value
dashboard = FairlearnDashboard(sensitive_features=A_test,
sensitive_feature_names=protected_features,
y_true=Y_test,
y_pred=dashboard_predicted)
return dashboard
def fairness_dashboard(A_test, protected_features, y_true, y_pred):
"""
Displays an interactive dashboards that shows fairness metrics given the data,
protected features and the train-test targets.
Parameters:
A_test (pandas dataframe): Input data to be analyzed for fairness metrics.
protected_features (list): Features or columns we need to account for potential biases.
y_true (list): List of ground truths
y_pred (list): List of predictions
Returns:
Interactive dashboard for fairness metrics.
"""
FairlearnDashboard(sensitive_features=A_test,
sensitive_feature_names=protected_features,
y_true=y_true.values.tolist(),
y_pred=[y_pred.tolist()])
# To show the effect of Fairlearn we will first train a standard ML predictor
# that does not incorporate fairness.
# For speed of demonstration, we use the simple
# :class:`sklearn.linear_models.LogisticRegression` class:
unmitigated_predictor = LogisticRegression(solver='liblinear',
fit_intercept=True)
unmitigated_predictor.fit(X_train, Y_train)
# %%
# We can load this predictor into the Fairness dashboard, and assess its fairness:
FairlearnDashboard(
sensitive_features=A_test,
sensitive_feature_names=['sex'],
y_true=Y_test,
y_pred={"unmitigated": unmitigated_predictor.predict(X_test)})
# %%
# Looking at the disparity in accuracy, we see that males have an error
# about three times greater than the females.
# More interesting is the disparity in opportunity - males are offered loans at
# three times the rate of females.
#
# Despite the fact that we removed the feature from the training data, our
# predictor still discriminates based on sex.
# This demonstrates that simply ignoring a sensitive feature when fitting a
# predictor rarely eliminates unfairness.
# There will generally be enough other features correlated with the removed
# feature to lead to disparate impact.
y_pred,
sensitive_features=sex)
print("group_summary", result1)
result2 = metrics.selection_rate_group_summary(y_true,
y_pred,
sensitive_features=sex)
print("selection_rate_group_summary", result2)
# FairlearnDashboard(sensitive_features=sex,
# sensitive_feature_names=['sex'],
# y_true=y_true,
# y_pred={"initial model": y_pred})
np.random.seed(0)
constraint = DemographicParity()
classifier = DecisionTreeClassifier()
mitigator = ExponentiatedGradient(classifier, constraint)
#print("constructing mitigator")
mitigator.fit(X, y_true, sensitive_features=sex)
y_pred_mitigated = mitigator.predict(X)
result2_mitigated = metrics.selection_rate_group_summary(
y_true, y_pred_mitigated, sensitive_features=sex)
print("selection_rate_group_summary mitigated", result2_mitigated)
FairlearnDashboard(sensitive_features=sex,
sensitive_feature_names=['sex'],
y_true=y_true,
y_pred={
"initial model": y_pred,
"mitigated model": y_pred_mitigated
})
#FairlearnDashboard(sensitive_features=sex, sensitive_feature_names=['sex'],y_true=y_true,y_pred={"initial model": y_pred})