def test_standard_checks(test_fn):
trf = FairClassifier(
covariance_threshold=None,
C=1,
penalty="none",
sensitive_cols=[0],
train_sensitive_cols=True,
)
test_fn(FairClassifier.__name__, trf)
def test_regularization(sensitive_classification_dataset):
"""Tests whether increasing regularization decreases the norm of the coefficient vector"""
X, y = sensitive_classification_dataset
prev_theta_norm = np.inf
for C in [1, 0.5, 0.2, 0.1]:
fair = FairClassifier(
covariance_threshold=None, sensitive_cols=["x1"], C=C
).fit(X, y)
theta_norm = np.abs(np.sum(fair.coef_))
assert theta_norm < prev_theta_norm
prev_theta_norm = theta_norm
def test_deprecation():
with warnings.catch_warnings(record=True) as w:
# Cause all warnings to always be triggered.
warnings.simplefilter("always")
# Trigger a warning.
FairClassifier(
covariance_threshold=1,
sensitive_cols=["x1"],
penalty="none",
train_sensitive_cols=False,
)
assert issubclass(w[-1].category, DeprecationWarning)
def test_fairness(sensitive_classification_dataset):
"""tests whether fairness (measured by p percent score) increases as we decrease the covariance threshold"""
X, y = sensitive_classification_dataset
scorer = p_percent_score("x1")
prev_fairness = -np.inf
for cov_threshold in [None, 10, 0.5, 0.1]:
fair = FairClassifier(
covariance_threshold=cov_threshold,
sensitive_cols=["x1"],
penalty="none",
train_sensitive_cols=False,
).fit(X, y)
fairness = scorer(fair, X, y)
assert fairness >= prev_fairness
prev_fairness = fairness
def _test_same(dataset):
X, y = dataset
if X.shape[1] == 1:
# If we only have one column (which is also the sensitive one) we can't fit
return True
sensitive_cols = [0]
X_without_sens = np.delete(X, sensitive_cols, axis=1)
lr = LogisticRegression(penalty="none", solver="lbfgs")
fair = FairClassifier(
covariance_threshold=None, sensitive_cols=sensitive_cols, penalty="none"
)
try:
fair.fit(X, y)
except SolverError:
pass
else:
lr.fit(X_without_sens, y)
normal_pred = lr.predict_proba(X_without_sens)
fair_pred = fair.predict_proba(X)
np.testing.assert_almost_equal(normal_pred, fair_pred, decimal=2)
assert np.sum(lr.predict(X_without_sens) != fair.predict(X)) / len(X) < 0.01