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",
multi_class="ovr",
dual=False,
tol=1e-4,
C=1.0,
fit_intercept=True,
intercept_scaling=1,
class_weight=None,
random_state=None,
max_iter=100,
verbose=0,
warm_start=False,
n_jobs=None,
l1_ratio=None,
)
fair = EqualOpportunityClassifier(
covariance_threshold=None,
sensitive_cols=sensitive_cols,
penalty="none",
positive_target=True,
)
fair.fit(X, y)
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
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 = EqualOpportunityClassifier(
covariance_threshold=None,
sensitive_cols=sensitive_cols,
penalty="none",
positive_target=True,
)
fair.fit(X, y)
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
def objective(self, trial):
C = trial.suggest_loguniform('C', 1e-5, 1e5)
c = trial.suggest_loguniform('c', 1e-5, 1e5)
#print(c, C)
try:
#if 1==1:
if self.base_model=='equal':
model = EqualOpportunityClassifier(sensitive_cols=self.fair_feat, positive_target=True, covariance_threshold=c, C=C, max_iter=10**3)
model.fit(self.X_train, self.y_train)
elif self.base_model=='demographic':
model = DemographicParityClassifier(sensitive_cols=self.fair_feat, covariance_threshold=c, C=C, max_iter=10**3)
model.fit(self.X_train, self.y_train)
elif self.base_model=='minimax':
a_train = self.X_train[self.fair_feat].copy().astype('int')
a_val = self.X_val[self.fair_feat].copy().astype('int')
a_train[a_train==0] = -1
a_val[a_val==0] = -1
model = SKLearn_Weighted_LLR(self.X_train.values, self.y_train.values,
a_train.values, self.X_val.values,
self.y_val.values, a_val.values,
C_reg=C)
mua_ini = np.ones(a_val.max() + 1)
mua_ini /= mua_ini.sum()
results = APSTAR(model, mua_ini, niter=200, max_patience=200, Kini=1,
Kmin=20, alpha=0.5, verbose=False)
mu_best_list = results['mu_best_list']
mu_best = mu_best_list[-1]
model.weighted_fit(self.X_train.values, self.y_train.values, a_train.values, mu_best)
else:
raise('Incorrect base_model.')
y_pred = model.predict(self.X_val)
except:
return float('inf')
if (sklearn.metrics.accuracy_score(self.y_val, y_pred)==0 or
equal_opportunity_score(sensitive_column=self.fair_feat)(model, self.X_val, self.y_val)==0 or
p_percent_score(sensitive_column=self.fair_feat)(model, self.X_val))==0:
return float('inf')
if self.metric=='accuracy':
perf = sklearn.metrics.accuracy_score(self.y_val, y_pred)
elif self.metric=='equal_opportunity':
perf = equal_opportunity_score(sensitive_column=self.fair_feat)(model, self.X_val, self.y_val)
elif self.metric=='p_percent':
perf = p_percent_score(sensitive_column=self.fair_feat)(model, self.X_val)
elif self.metric=='c_variation':
perf = 1/coefficient_of_variation(model, self.X_val, self.y_val)
if perf>self.best_perf:
self.best_perf = perf
self.best_model = model
return 1/perf if perf!=0 else float('inf')