def decision_function(self, X):
return self.project(X)
def predict(self, X):
return np.sign(self.project(X))
def score(self, X_test, y_test):
predict = self.predict(X_test)
acc = accuracy_score(y_test, predict)
return acc
if __name__ == "__main__":
# Load Adult dataset (a smaller version!)
dataset_train, dataset_test = load_adult(smaller=True)
sensible_feature = 9 # GENDER
sensible_feature_values = sorted(
list(set(dataset_train.data[:, sensible_feature])))
print('Different values of the sensible feature', sensible_feature, ':',
sensible_feature_values)
ntrain = len(dataset_train.target)
# Standard SVM - Train an SVM using the training set
print('Grid search for SVM...')
grid_search_complete = 1
if grid_search_complete:
param_grid = [{
'C': [0.1, 1, 10.0],
'gamma': [0.1, 0.01],
'kernel': ['rbf']
#self.coef_ = self.model.coef_
#self.intercept_ = self.model.intercept_
if __name__ == "__main__":
experiment_number = 0
if experiment_number == 0:
dataset_train = load_binary_diabetes_uci()
dataset_test = load_binary_diabetes_uci()
sensible_feature = 1 # sex
elif experiment_number == 1:
dataset_train = load_heart_uci()
dataset_test = load_heart_uci()
sensible_feature = 1 # sex
elif experiment_number == 2:
dataset_train, dataset_test = load_adult(smaller=False)
sensible_feature = 9 # sex
print('Different values of the sensible feature', sensible_feature,
':', set(dataset_train.data[:, sensible_feature]))
elif experiment_number == 3:
dataset_train, dataset_test = load_adult_race(smaller=False)
sensible_feature = 8 # race
print('Different values of the sensible feature', sensible_feature,
':', set(dataset_train.data[:, sensible_feature]))
if experiment_number in [0, 1]:
# % for train
ntrain = 5 * len(dataset_train.target) // 10
dataset_train.data = dataset_train.data[:ntrain, :]
dataset_train.target = dataset_train.target[:ntrain]
dataset_test.data = dataset_test.data[ntrain:, :]
def experiment(dataset,
frac,
eval_objective,
eps,
rho_list,
rho,
eps_list,
criteria,
classifier,
trials,
include_sensible,
filename,
learner_name='lsq',
mode='four',
verbose=False):
'''
dataset: one of ['compas', 'bank', 'adult', 'law', 'german']. Default is 'compas'.
frac: real number in interval [0, 1]. The fraction of the data points in chosen dataset to use.
eval_objective: ['test_tau', 'test_rho_est_err']. 'test_tau' runs experiments between tau and error/fairness violation. 'test_rho_est_err' runs experiments between the estimated rho and error/fairness violation.
eps: a number specifying the wanted fairness level. Valid when eval_objective='test_rho_est_err'.
rho_list: a list of (rho_plut, rho_minus) pairs. Valid when eval_objective='test_rho_est_err'.
rho: [a, b] where a, b in interval [0,0.5].
eps_list: a list of non-negative real numbers. Valid when eval_objective='test_eps'.
criteria: one of ['DP','EO']
classifier: one of ['Agarwal', 'Zafar']. Agarwal is the default.
trials: the number of trials to run.
include_sensible: boolean. If to include sensitive attribute as a feature for optimizing the oroginal loss. This is used only for debugging purpose. It is hard-coded to be False now.
filename: the file name to store the log of experiment(s).
learner_name: ['lsq', 'LR', 'SVM']. SVM is the slowest. lsq does not work for law school dataset but it works reasonally well on all other datasets.
mode: ['four']. Currently, we only support four. Valid when eval_objective='test_eps'.
verbose: boolean. If print out info at each run.
'''
# We hard-code mode and classifier.
mode = 'four'
# classifier
if classifier not in ['Agarwal', 'Zafar']:
classifier = 'Agarwal'
# We hard-code include_sensible to False.
include_sensible = False
sensible_name = None
sensible_feature = None
learner = None
print('input dataset:', dataset)
if dataset == 'adult':
datamat = load_adult(frac)
sensible_name = 'gender'
sensible_feature = 9
elif dataset == 'law':
datamat = load_law(frac)
sensible_name = 'racetxt'
sensible_feature = 9
# lsq does not work for law
learner_name = 'LR'
elif dataset == 'german':
datamat = load_german(frac)
sensible_name = 'Foreign'
sensible_feature = 21
elif dataset == 'bank':
datamat = load_bank(frac)
sensible_name = 'Middle_Aged'
sensible_feature = 7
else:
datamat = load_compas(frac)
sensible_name = 'race'
sensible_feature = 4
if learner_name == 'LR':
learner = LR()
elif learner_name == 'SVM':
learner = SVM()
else:
learner = LeastSquaresLearner()
print('eval_objective', eval_objective)
print('learner_name:', learner_name)
if eval_objective == 'test_rho_est_err':
eps_list = [eps for _ in range(len(rho_list))]
if criteria == 'EO':
tests = [{"cons_class": moments.EO, "eps": eps} for eps in eps_list]
else:
tests = [{"cons_class": moments.DP, "eps": eps} for eps in eps_list]
if eval_objective == 'test_rho_est_err':
all_data = _experiment_est_error(datamat, tests, rho, rho_list, trials,
sensible_name, sensible_feature,
criteria, classifier,
include_sensible, learner, mode,
verbose)
_save_all_data(filename, all_data, rho_list)
else:
all_data = _experiment(datamat, tests, rho, trials, sensible_name,
sensible_feature, criteria, classifier,
include_sensible, learner, mode, verbose)
_save_all_data(filename, all_data, eps_list)
return all_data
