def thresholdOptimizer(X_train, Y_train, A_train, model, constraint):
"""
Parameters:
y_train: input data for training the model
X_train: list of ground truths
constraints: either "demographic_parity" or "equalized_odds"
Returns the predictions of the optimized model
"""
postprocess_est = ThresholdOptimizer(estimator=model,
constraints=constraint)
# Balanced data set is obtained by sampling the same number of points from the majority class (Y=0)
# as there are points in the minority class (Y=1)
Y_train = pd.Series(Y_train)
balanced_idx1 = X_train[[Y_train == 1]].index
pp_train_idx = balanced_idx1.union(Y_train[Y_train == 0].sample(
n=balanced_idx1.size, random_state=1234).index)
X_train_balanced = X_train.loc[pp_train_idx, :]
Y_train_balanced = Y_train.loc[pp_train_idx]
A_train_balanced = A_train.loc[pp_train_idx]
postprocess_est.fit(X_train_balanced,
Y_train_balanced,
sensitive_features=A_train_balanced)
postprocess_preds = postprocess_est.predict(X_test,
sensitive_features=A_test)
return postprocess_preds
def fit(self, _X, _Y, _classifier_name="logistic", _predictor="hard"):
my_erm_classifier = erm_classifier(self.train_X, self.train_Y)
my_erm_classifier.fit(self.train_X, self.train_Y, classifier_name=_classifier_name)
self.model = ThresholdOptimizer(estimator=my_erm_classifier, \
constraints="demographic_parity", prefit=True)
self.model.fit(self.train_X, self.train_Y, \
sensitive_features=self.sensitive_train, _predictor=_predictor)
def test_threshold_optimization_equalized_odds_e2e(
sensitive_features, sensitive_feature_names, expected_positive_p0,
expected_positive_p1, expected_negative_p0, expected_negative_p1,
X_transform, y_transform, sensitive_features_transform):
X = X_transform(_format_as_list_of_lists(sensitive_features))
y = y_transform(labels_ex)
sensitive_features_ = sensitive_features_transform(sensitive_features)
adjusted_predictor = ThresholdOptimizer(
unconstrained_predictor=ExamplePredictor(), constraints=EQUALIZED_ODDS)
adjusted_predictor.fit(X, y, sensitive_features=sensitive_features_)
predictions = adjusted_predictor._pmf_predict(
X, sensitive_features=sensitive_features_)
# assert equalized odds
for a in sensitive_feature_names:
positive_indices = (np.array(sensitive_features) == a) * \
(np.array(labels_ex) == 1)
negative_indices = (np.array(sensitive_features) == a) * \
(np.array(labels_ex) == 0)
average_probs_positive_indices = np.average(
predictions[positive_indices], axis=0)
average_probs_negative_indices = np.average(
predictions[negative_indices], axis=0)
assert np.isclose(average_probs_positive_indices[0],
expected_positive_p0)
assert np.isclose(average_probs_positive_indices[1],
expected_positive_p1)
assert np.isclose(average_probs_negative_indices[0],
expected_negative_p0)
assert np.isclose(average_probs_negative_indices[1],
expected_negative_p1)
def test_threshold_optimization_equalized_odds_e2e(data_X_y_sf):
adjusted_predictor = ThresholdOptimizer(
estimator=ExamplePredictor(scores_ex), constraints=EQUALIZED_ODDS)
adjusted_predictor.fit(data_X_y_sf.X,
data_X_y_sf.y,
sensitive_features=data_X_y_sf.sensitive_features)
predictions = adjusted_predictor._pmf_predict(
data_X_y_sf.X, sensitive_features=data_X_y_sf.sensitive_features)
expected_ps = _expected_ps_equalized_odds[data_X_y_sf.example_name]
mapped_sensitive_features = _map_into_single_column(
data_X_y_sf.sensitive_features)
# assert equalized odds
for a in data_X_y_sf.feature_names:
pos_indices = (mapped_sensitive_features == a) * (labels_ex == 1)
neg_indices = (mapped_sensitive_features == a) * (labels_ex == 0)
average_probs_positive_indices = np.average(predictions[pos_indices],
axis=0)
average_probs_negative_indices = np.average(predictions[neg_indices],
axis=0)
assert np.isclose(average_probs_positive_indices[0],
expected_ps[_POS_P0])
assert np.isclose(average_probs_positive_indices[1],
expected_ps[_POS_P1])
assert np.isclose(average_probs_negative_indices[0],
expected_ps[_NEG_P0])
assert np.isclose(average_probs_negative_indices[1],
expected_ps[_NEG_P1])
def fit(self):
self.erm_classifier.fit(self.train_X, self.train_Y)
self.model = ThresholdOptimizer(estimator=self.erm_classifier,
constraints=self.metric,
prefit=True)
self.model.fit(self.train_X,
self.train_Y,
sensitive_features=self.sensitive_train)
def run_thresholdoptimizer_classification(estimator):
"""Run classification test with ThresholdOptimizer."""
X, Y, A = fetch_adult()
to = ThresholdOptimizer(estimator=estimator, prefit=False)
to.fit(X, Y, sensitive_features=A)
results = to.predict(X, sensitive_features=A)
assert results is not None
def _fit_and_plot(constraints, plotting_data):
adjusted_predictor = ThresholdOptimizer(
estimator=ExamplePredictor(scores_ex), constraints=constraints)
adjusted_predictor.fit(plotting_data.X,
plotting_data.y,
sensitive_features=plotting_data.sensitive_features)
fig, (ax) = plt.subplots(1, 1)
plot_threshold_optimizer(adjusted_predictor, ax=ax, show_plot=False)
return fig
def test_threshold_optimization_degenerate_labels(data_X_sf, y_transform, constraints):
y = y_transform(degenerate_labels_ex)
adjusted_predictor = ThresholdOptimizer(estimator=ExamplePredictor(scores_ex),
constraints=constraints,
predict_method='predict')
feature_name = _degenerate_labels_feature_name[data_X_sf.example_name]
with pytest.raises(ValueError, match=DEGENERATE_LABELS_ERROR_MESSAGE.format(feature_name)):
adjusted_predictor.fit(data_X_sf.X, y,
sensitive_features=data_X_sf.sensitive_features)
def test_predict_output_0_or_1(data_X_y_sf, constraints):
adjusted_predictor = ThresholdOptimizer(estimator=ExamplePredictor(scores_ex),
constraints=constraints,
predict_method='predict')
adjusted_predictor.fit(data_X_y_sf.X, data_X_y_sf.y,
sensitive_features=data_X_y_sf.sensitive_features)
predictions = adjusted_predictor.predict(
data_X_y_sf.X, sensitive_features=data_X_y_sf.sensitive_features)
for prediction in predictions:
assert prediction in [0, 1]
def test_threshold_optimization_non_binary_labels(data_X_y_sf, constraints):
non_binary_y = deepcopy(data_X_y_sf.y)
non_binary_y[0] = 2
adjusted_predictor = ThresholdOptimizer(estimator=ExamplePredictor(scores_ex),
constraints=constraints,
predict_method='predict')
with pytest.raises(ValueError, match=_LABELS_NOT_0_1_ERROR_MESSAGE):
adjusted_predictor.fit(data_X_y_sf.X, non_binary_y,
sensitive_features=data_X_y_sf.sensitive_features)
def test_threshold_optimization_degenerate_labels(X_transform, y_transform,
sensitive_features_transform,
constraints):
X = X_transform(_format_as_list_of_lists(sensitive_features_ex1))
y = y_transform(degenerate_labels_ex)
sensitive_features = sensitive_features_transform(sensitive_features_ex1)
adjusted_predictor = ThresholdOptimizer(
unconstrained_predictor=ExamplePredictor(), constraints=constraints)
with pytest.raises(ValueError,
match=DEGENERATE_LABELS_ERROR_MESSAGE.format('A')):
adjusted_predictor.fit(X, y, sensitive_features=sensitive_features)
def test_inconsistent_input_data_types(X, y, sensitive_features, constraints):
adjusted_predictor = ThresholdOptimizer(
unconstrained_predictor=ExamplePredictor(), constraints=constraints)
error_message = INPUT_DATA_FORMAT_ERROR_MESSAGE.format(
type(X).__name__,
type(y).__name__,
type(sensitive_features).__name__)
if X is None or y is None and sensitive_features is None:
with pytest.raises(TypeError) as exception:
adjusted_predictor.fit(X, y, sensitive_features=sensitive_features)
assert str(exception.value) == error_message
def test_predict_output_0_or_1(sensitive_features, sensitive_feature_names,
X_transform, y_transform,
sensitive_features_transform, constraints):
X = X_transform(_format_as_list_of_lists(sensitive_features))
y = y_transform(labels_ex)
sensitive_features_ = sensitive_features_transform(sensitive_features)
adjusted_predictor = ThresholdOptimizer(
unconstrained_predictor=ExamplePredictor(), constraints=constraints)
adjusted_predictor.fit(X, y, sensitive_features=sensitive_features_)
predictions = adjusted_predictor.predict(
X, sensitive_features=sensitive_features_)
for prediction in predictions:
assert prediction in [0, 1]
def test_predict_different_argument_lengths(data_X_y_sf, constraints):
adjusted_predictor = ThresholdOptimizer(
estimator=ExamplePredictor(scores_ex),
constraints=constraints,
predict_method="predict",
)
adjusted_predictor.fit(data_X_y_sf.X,
data_X_y_sf.y,
sensitive_features=data_X_y_sf.sensitive_features)
with pytest.raises(
ValueError,
match="Found input variables with inconsistent numbers of samples"
):
adjusted_predictor.predict(
data_X_y_sf.X,
sensitive_features=data_X_y_sf.sensitive_features[:-1])
with pytest.raises(
ValueError,
match="Found input variables with inconsistent numbers of samples"
):
adjusted_predictor.predict(
data_X_y_sf.X[:-1],
sensitive_features=data_X_y_sf.sensitive_features)
def test_threshold_optimization_different_input_lengths(data_X_y_sf, constraints):
n = len(X_ex)
expected_exception_messages = {
"inconsistent": 'Found input variables with inconsistent numbers of samples',
"empty": 'Found array with 0 sample'
}
for permutation in [(0, 1), (1, 0)]:
with pytest.raises(ValueError, match=expected_exception_messages['inconsistent']
.format("X, sensitive_features, and y")):
adjusted_predictor = ThresholdOptimizer(
estimator=ExamplePredictor(scores_ex),
constraints=constraints,
predict_method='predict')
adjusted_predictor.fit(data_X_y_sf.X[:n - permutation[0]],
data_X_y_sf.y[:n - permutation[1]],
sensitive_features=data_X_y_sf.sensitive_features)
# try providing empty lists in all combinations
for permutation in [(0, n, 'inconsistent'), (n, 0, 'empty')]:
adjusted_predictor = ThresholdOptimizer(
estimator=ExamplePredictor(scores_ex),
constraints=constraints,
predict_method='predict')
with pytest.raises(ValueError, match=expected_exception_messages[permutation[2]]):
adjusted_predictor.fit(data_X_y_sf.X[:n - permutation[0]],
data_X_y_sf.y[:n - permutation[1]],
sensitive_features=data_X_y_sf.sensitive_features)
def test_predict_different_argument_lengths(sensitive_features,
sensitive_feature_names,
X_transform, y_transform,
sensitive_features_transform,
constraints):
X = X_transform(_format_as_list_of_lists(sensitive_features))
y = y_transform(labels_ex)
sensitive_features_ = sensitive_features_transform(sensitive_features)
adjusted_predictor = ThresholdOptimizer(
unconstrained_predictor=ExamplePredictor(), constraints=constraints)
adjusted_predictor.fit(X, y, sensitive_features=sensitive_features_)
with pytest.raises(ValueError,
match=DIFFERENT_INPUT_LENGTH_ERROR_MESSAGE.format(
"X and sensitive_features")):
adjusted_predictor.predict(
X,
sensitive_features=sensitive_features_transform(
sensitive_features[:-1]))
with pytest.raises(ValueError,
match=DIFFERENT_INPUT_LENGTH_ERROR_MESSAGE.format(
"X and sensitive_features")):
adjusted_predictor.predict(X_transform(
_format_as_list_of_lists(sensitive_features))[:-1],
sensitive_features=sensitive_features_)
def test_threshold_optimization_different_input_lengths(
X_transform, y_transform, sensitive_features_transform, constraints):
n = len(sensitive_features_ex1)
for permutation in [(0, 1), (1, 0)]:
with pytest.raises(ValueError,
match=DIFFERENT_INPUT_LENGTH_ERROR_MESSAGE.format(
"X, sensitive_features, and y")):
X = X_transform(
_format_as_list_of_lists(sensitive_features_ex1)
[:n - permutation[0]])
y = y_transform(labels_ex[:n - permutation[1]])
sensitive_features = sensitive_features_transform(
sensitive_features_ex1)
adjusted_predictor = ThresholdOptimizer(
unconstrained_predictor=ExamplePredictor(),
constraints=constraints)
adjusted_predictor.fit(X, y, sensitive_features=sensitive_features)
# try providing empty lists in all combinations
for permutation in [(0, n), (n, 0)]:
X = X_transform(
_format_as_list_of_lists(sensitive_features_ex1)[:n -
permutation[0]])
y = y_transform(labels_ex[:n - permutation[1]])
sensitive_features = sensitive_features_transform(
sensitive_features_ex1)
adjusted_predictor = ThresholdOptimizer(
unconstrained_predictor=ExamplePredictor(),
constraints=constraints)
with pytest.raises(ValueError, match=EMPTY_INPUT_ERROR_MESSAGE):
adjusted_predictor.fit(X, y, sensitive_features=sensitive_features)
def test_no_matplotlib(constraints):
n_samples = 50
n_features = 50
n_sensitive_feature_values = 2
n_classes = 2
threshold_optimizer = ThresholdOptimizer(unconstrained_predictor=FakePredictor(),
constraints=constraints,
plot=True)
with pytest.raises(RuntimeError) as exc:
threshold_optimizer.fit(X=np.random.random((n_samples, n_features)),
y=np.random.randint(n_classes, size=n_samples),
sensitive_features=np.random.randint(n_sensitive_feature_values,
size=n_samples))
assert str(exc.value) == _MATPLOTLIB_IMPORT_ERROR_MESSAGE
def test_threshold_optimization_non_binary_labels(X_transform, y_transform,
sensitive_features_transform,
constraints):
non_binary_labels = copy.deepcopy(labels_ex)
non_binary_labels[0] = 2
X = X_transform(_format_as_list_of_lists(sensitive_features_ex1))
y = y_transform(non_binary_labels)
sensitive_features = sensitive_features_transform(sensitive_features_ex1)
adjusted_predictor = ThresholdOptimizer(
unconstrained_predictor=ExamplePredictor(), constraints=constraints)
with pytest.raises(ValueError, match=NON_BINARY_LABELS_ERROR_MESSAGE):
adjusted_predictor.fit(X, y, sensitive_features=sensitive_features)
def test_constraints_not_supported():
with pytest.raises(ValueError, match=NOT_SUPPORTED_CONSTRAINTS_ERROR_MESSAGE):
ThresholdOptimizer(estimator=ExamplePredictor(scores_ex),
constraints="UnsupportedConstraints",
predict_method='predict').fit(
X_ex, labels_ex,
sensitive_features=sensitive_features_ex1
)
def test_random_state_threshold_optimizer():
"""Test that the random_state argument works as expected.
This test case reproduces the problem reported in issue 588 if the
random_state does not work as intended within ThresholdOptimizer.
https://github.com/fairlearn/fairlearn/issues/588
"""
X_train, X_test, y_train, y_test, race_train, race_test = _get_test_data()
# Train a simple logistic regression model
lr = LogisticRegression(max_iter=1000, random_state=0)
lr.fit(X_train, y_train)
# Train threshold optimizer
to = ThresholdOptimizer(estimator=lr,
constraints='equalized_odds',
grid_size=1000)
to.fit(X_train, y_train, sensitive_features=race_train)
# score groups
y_pred_test = to.predict(X_test,
sensitive_features=race_test,
random_state=0)
for _ in range(100):
assert (y_pred_test == to.predict(X_test,
sensitive_features=race_test,
random_state=0)).all()
assert (y_pred_test != to.predict(
X_test, sensitive_features=race_test, random_state=1)).any()
def test_threshold_optimization_demographic_parity_e2e(data_X_y_sf):
adjusted_predictor = ThresholdOptimizer(
estimator=ExamplePredictor(scores_ex), constraints=DEMOGRAPHIC_PARITY)
adjusted_predictor.fit(data_X_y_sf.X,
data_X_y_sf.y,
sensitive_features=data_X_y_sf.sensitive_features)
predictions = adjusted_predictor._pmf_predict(
data_X_y_sf.X, sensitive_features=data_X_y_sf.sensitive_features)
expected_ps = _expected_ps_demographic_parity[data_X_y_sf.example_name]
# assert demographic parity
for sensitive_feature_name in data_X_y_sf.feature_names:
average_probs = np.average(predictions[_map_into_single_column(
data_X_y_sf.sensitive_features) == sensitive_feature_name],
axis=0)
assert np.isclose(average_probs[0], expected_ps[_P0])
assert np.isclose(average_probs[1], expected_ps[_P1])
def test_both_predictor_and_estimator_error(constraints):
with pytest.raises(ValueError,
match=EITHER_PREDICTOR_OR_ESTIMATOR_ERROR_MESSAGE):
ThresholdOptimizer(unconstrained_predictor=ExamplePredictor(scores_ex),
estimator=ExampleEstimator(),
constraints=constraints).fit(
X_ex,
labels_ex,
sensitive_features=sensitive_features_ex1)
def test_predict_before_fit_error(X_transform, sensitive_features_transform,
predict_method_name, constraints):
X = X_transform(_format_as_list_of_lists(sensitive_features_ex1))
sensitive_features = sensitive_features_transform(sensitive_features_ex1)
adjusted_predictor = ThresholdOptimizer(
unconstrained_predictor=ExamplePredictor(), constraints=constraints)
with pytest.raises(ValueError, match=PREDICT_BEFORE_FIT_ERROR_MESSAGE):
getattr(adjusted_predictor,
predict_method_name)(X, sensitive_features=sensitive_features)
class demographic_parity_classifier(base_binary_classifier):
def fit(self, _X, _Y, _classifier_name="logistic", _predictor="hard"):
my_erm_classifier = erm_classifier(self.train_X, self.train_Y)
my_erm_classifier.fit(self.train_X, self.train_Y, classifier_name=_classifier_name)
self.model = ThresholdOptimizer(estimator=my_erm_classifier, \
constraints="demographic_parity", prefit=True)
self.model.fit(self.train_X, self.train_Y, \
sensitive_features=self.sensitive_train, _predictor=_predictor)
def predict(self, x_samples, sensitive_features):
y_samples = self.model.predict(x_samples, sensitive_features=sensitive_features)
return y_samples
def get_accuracy(self, X, y_true, sensitive_features):
y_pred = self.predict(X, sensitive_features)
return 1 - np.sum(np.power(y_pred - y_true, 2))/len(y_true)
def predict_proba(self, x_samples, sensitive_features):
y_samples = self.model._pmf_predict(x_samples, sensitive_features=sensitive_features)
return y_samples
def test_predict_method(predict_method):
class Dummy(BaseEstimator, ClassifierMixin):
def fit(self, X, y):
return self
def predict(self, X):
raise Exception("predict")
def predict_proba(self, X):
raise Exception("predict_proba")
def decision_function(self, X):
raise Exception("decision_function")
X, y = make_classification()
sensitive_feature = np.random.randint(0, 2, len(y))
clf = ThresholdOptimizer(estimator=Dummy(), predict_method=predict_method)
exception = "predict_proba" if predict_method == "auto" else predict_method
with pytest.raises(Exception, match=exception):
clf.fit(X, y, sensitive_features=sensitive_feature)
def test_predict_before_fit_error(X_transform, sensitive_features_transform,
predict_method_name, constraints):
X = X_transform(sensitive_features_ex1)
sensitive_features = sensitive_features_transform(sensitive_features_ex1)
adjusted_predictor = ThresholdOptimizer(
unconstrained_predictor=ExamplePredictor(scores_ex),
constraints=constraints)
with pytest.raises(ValueError, match='instance is not fitted yet'):
getattr(adjusted_predictor,
predict_method_name)(X, sensitive_features=sensitive_features)
def run_thresholdoptimizer_classification(estimator):
"""Run classification test with ThresholdOptimizer."""
X_train, Y_train, A_train, X_test, Y_test, A_test = fetch_adult()
unmitigated = copy.deepcopy(estimator)
unmitigated.fit(X_train, Y_train)
unmitigated_predictions = unmitigated.predict(X_test)
to = ThresholdOptimizer(estimator=estimator,
prefit=False,
predict_method='predict')
to.fit(X_train, Y_train, sensitive_features=A_train)
mitigated_predictions = to.predict(X_test, sensitive_features=A_test)
dp_diff_unmitigated = demographic_parity_difference(
Y_test, unmitigated_predictions, sensitive_features=A_test)
dp_diff_mitigated = demographic_parity_difference(
Y_test, mitigated_predictions, sensitive_features=A_test)
assert dp_diff_mitigated <= dp_diff_unmitigated
def test_threshold_optimization_demographic_parity_e2e(
sensitive_features, sensitive_feature_names, expected_p0, expected_p1,
X_transform, y_transform, sensitive_features_transform):
X = X_transform(_format_as_list_of_lists(sensitive_features))
y = y_transform(labels_ex)
sensitive_features_ = sensitive_features_transform(sensitive_features)
adjusted_predictor = ThresholdOptimizer(
unconstrained_predictor=ExamplePredictor(),
constraints=DEMOGRAPHIC_PARITY)
adjusted_predictor.fit(X, y, sensitive_features=sensitive_features_)
predictions = adjusted_predictor._pmf_predict(
X, sensitive_features=sensitive_features_)
# assert demographic parity
for sensitive_feature_name in sensitive_feature_names:
average_probs = np.average(predictions[np.array(sensitive_features) ==
sensitive_feature_name],
axis=0)
assert np.isclose(average_probs[0], expected_p0)
assert np.isclose(average_probs[1], expected_p1)
def test_constraints_objective_pairs(constraints, objective):
X = pd.Series(
[0, 1, 2, 3, 4, 0, 1, 2, 3]).to_frame()
sf = pd.Series(
[0, 0, 0, 0, 0, 1, 1, 1, 1])
y = pd.Series(
[1, 0, 1, 1, 1, 0, 1, 1, 1])
thr_optimizer = ThresholdOptimizer(
estimator=PassThroughPredictor(),
constraints=constraints,
objective=objective,
grid_size=20,
predict_method='predict')
expected = results[constraints+", "+objective]
if type(expected) is str:
with pytest.raises(ValueError) as error_info:
thr_optimizer.fit(X, y, sensitive_features=sf)
assert str(error_info.value).startswith(expected)
else:
thr_optimizer.fit(X, y, sensitive_features=sf)
res = thr_optimizer.interpolated_thresholder_.interpolation_dict
for key in [0, 1]:
assert res[key]['p0'] == pytest.approx(expected[key]['p0'], PREC)
assert res[key]['operation0']._operator == expected[key]['op0']
assert res[key]['operation0']._threshold == pytest.approx(expected[key]['thr0'], PREC)
assert res[key]['p1'] == pytest.approx(expected[key]['p1'], PREC)
assert res[key]['operation1']._operator == expected[key]['op1']
assert res[key]['operation1']._threshold == pytest.approx(expected[key]['thr1'], PREC)
if 'p_ignore' in expected[key]:
assert res[key]['p_ignore'] == pytest.approx(expected[key]['p_ignore'], PREC)
assert res[key]['prediction_constant'] == \
pytest.approx(expected[key]['prediction_constant'], PREC)
else:
assert 'p_ignore' not in res[key]
