def train_test_classifier(): w = fdm.train_model_disp_mist(x_train, y_train, x_control_train, loss_function, EPS, cons_params) train_score, test_score, cov_all_train, cov_all_test, s_attr_to_fp_fn_train, s_attr_to_fp_fn_test = fdm.get_clf_stats(w, x_train, y_train, x_control_train, x_test, y_test, x_control_test, sensitive_attrs) # accuracy and FPR are for the test because we need of for plotting return w, test_score, s_attr_to_fp_fn_test
def train_test_classifier():
w = fdm.train_model_disp_mist(x_train, y_train, x_control_train, loss_function, EPS, cons_params)
train_score, test_score, cov_all_train, cov_all_test, s_attr_to_fp_fn_train, s_attr_to_fp_fn_test = fdm.get_clf_stats(
w, x_train, y_train, x_control_train, x_test, y_test, x_control_test, sensitive_attrs)
# accuracy and FPR are for the test because we need of for plotting
# the covariance is for train, because we need it for setting the thresholds
return w, train_score, test_score, s_attr_to_fp_fn_test, s_attr_to_fp_fn_train, cov_all_train
def train_classifier(x, y, control, cons_params):
loss_function = "logreg"
EPS = 0.0001
w = fdm.train_model_disp_mist(
x, y, control, loss_function,
EPS,
cons_params
)
return w
def train_classifier(x, y, control, cons_params, eps):
loss_function = "logreg" # only logistic regression is implemented
w = fdm.train_model_disp_mist(x, y, control, loss_function, eps,
cons_params)
return w
def _run_test_Zafar(test, data, sensible_name, learner, creteria):
'''
Invoking Zafar's algorithm.
'''
dataX, dataY, dataA, dataX_train, dataY_train, dataA_train, dataX_test, dataY_test, dataA_test = data
x, y, x_control, x_train, y_train, x_control_train, x_test, y_test, x_control_test = _convert_data_format_Zafar(
data, sensible_name)
w = None
if creteria == 'EO':
loss_function = "logreg" # perform the experiments with logistic regression
EPS = 1e-6
cons_type = 1 # FPR constraint -- just change the cons_type, the rest of parameters should stay the same
tau = 5.0
mu = 1.2
sensitive_attrs_to_cov_thresh = {
sensible_name: {
0: {
0: 0,
1: test['eps']
},
1: {
0: 0,
1: test['eps']
},
2: {
0: 0,
1: test['eps']
}
}
} # zero covariance threshold, means try to get the fairest solution
cons_params = {
"cons_type": cons_type,
"tau": tau,
"mu": mu,
"sensitive_attrs_to_cov_thresh": sensitive_attrs_to_cov_thresh
}
w = fdm.train_model_disp_mist(x, y, x_control, loss_function, EPS,
cons_params)
else:
apply_fairness_constraints = 1 # set this flag to one since we want to optimize accuracy subject to fairness constraints
apply_accuracy_constraint = 0
sep_constraint = 0
loss_function = lf._logistic_loss
sensitive_attrs = [sensible_name]
# print('eps:',test['eps'])
sensitive_attrs_to_cov_thresh = {sensible_name: test['eps']}
gamma = None
w = ut.train_model(x, y, x_control, loss_function,
apply_fairness_constraints,
apply_accuracy_constraint, sep_constraint,
sensitive_attrs, sensitive_attrs_to_cov_thresh,
gamma)
y_pred_train = np.sign(np.dot(x_train, w))
disp_train = fair_measure(y_pred_train, dataA_train, dataY_train, creteria)
train_score = accuracy_score(y_train, y_pred_train)
y_pred_test = np.sign(np.dot(x_test, w))
disp_test = fair_measure(y_pred_test, dataA_test, dataY_test, creteria)
test_score = accuracy_score(y_test, y_pred_test)
res = dict()
res["disp_train"] = disp_train
res["disp_test"] = disp_test
res["error_train"] = 1 - train_score
res["error_test"] = 1 - test_score
return res