def load_compas_dataset():
"""
Collect the aif360 preprocessed Compas Data Set.
Charge descriptions are removed.
:return: The Compas Dataset, split into training and test sets
"""
dataset = CompasDataset(
features_to_drop=['c_charge_desc'] # Drop charge description, as they unnecessarily overloads the dataset
)
ind = int(len(dataset.instance_names) * 0.8)
train, test = dataset.split([ind])
return train, test
def get_compass(sensitive_feature_name,
remove_z=False,
file_path="/home/btd26/datasets/compas/",
file_name="compas.npy",
**kwargs):
z_idx = get_z_idx(sensitive_feature_name, compas_sensitive_features_dict)
# load file
file_add = os.path.join(file_path, file_name)
if os.path.exists(file_add):
M = np.load("/home/btd26/datasets/compas/compas.npy")
else:
from aif360.datasets import CompasDataset
compas = CompasDataset()
M = np.concatenate([compas.features, compas.labels], axis=1)
np.save(file_add, M)
X = M[:, :-1]
y = M[:, -1]
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.2,
random_state=42)
Xtr, Xts, Ztr, Zts, ytr, yts = extract_z(X_test,
X_train,
y_test,
y_train,
z_idx,
remove_z=remove_z)
return Xtr, Xts, ytr, yts, Ztr, Zts
def load_dataset(name):
if name == 'Adult':
ds = AdultDataset()
elif name == 'German':
ds = GermanDataset()
elif name == 'Compas':
ds = CompasDataset()
return ds, name
def getSmallCompasDataset():
dataset = CompasDataset()
dataset_orig = load_preproc_data_compas(['sex'])
features = ['sex', 'race', 'age', 'priors_count', 'c_charge_degree']
domainArray = getSmallCompasDomain()
features.append(dataset_orig.label_names[0])
simpleDomain = Domain(features, domainArray)
labels = [y[0] for y in dataset_orig.labels]
simpleSamples = dataset_orig.features
simpleSamples = np.c_[simpleSamples, labels]
return simpleDomain, simpleSamples
def get_data(dataset_used, protected_attribute_used):
if dataset_used == "adult":
dataset_orig = AdultDataset()
if protected_attribute_used == 1:
privileged_groups = [{'sex': 1}]
unprivileged_groups = [{'sex': 0}]
else:
privileged_groups = [{'race': 1}]
unprivileged_groups = [{'race': 0}]
elif dataset_used == "german":
dataset_orig = load_preproc_data_german()
dataset_orig.labels -= 1
if protected_attribute_used == 1:
privileged_groups = [{'sex': 1}]
unprivileged_groups = [{'sex': 0}]
else:
privileged_groups = [{'age': 1}]
unprivileged_groups = [{'age': 0}]
elif dataset_used == "compas":
dataset_orig = CompasDataset()
if protected_attribute_used == 1:
privileged_groups = [{'sex': 1}]
unprivileged_groups = [{'sex': 0}]
else:
privileged_groups = [{'race': 1}]
unprivileged_groups = [{'race': 0}]
elif dataset_used == "bank":
dataset_orig = BankDataset()
if protected_attribute_used == 1:
privileged_groups = [{'age': 1}]
unprivileged_groups = [{'age': 0}]
else:
privileged_groups = [{'race': 1}]
unprivileged_groups = [{'race': 0}]
else:
raise ValueError(f"{dataset_used} is not an available dataset.")
dataset_orig_train, dataset_orig_vt = dataset_orig.split([0.6], shuffle=True, seed=101)
dataset_orig_valid, dataset_orig_test = dataset_orig_vt.split([0.5], shuffle=True, seed=101)
return dataset_orig_train, dataset_orig_valid, dataset_orig_test, privileged_groups, unprivileged_groups
def getLargeCompasDataset():
dataset = CompasDataset()
dataset_orig = load_preproc_data_compas(['sex'])
(head,types,records)=read_dataset()
records = np.array(records)
head,types = head[:-1], types[:-1]
records = np.delete(records, -1, axis=1)
def reorder(record):
record.pop(11)
record.append(record.pop(10))
return record
records = [reorder(list(record)) for record in records]
head = reorder(head)
records = np.array(records)
def getHotEncoding(index, n):
temp = [0]*n
temp[index] = 1
return tuple(temp)
races = list(set(records[:, 2]))
print (races)
fmo = list(set(records[:, 9]))
nrecords = []
domainArray = [set([]) for h in head]
for record in records:
temp = []
for j, (r, h) in enumerate(zip(record, head)):
if h == "sex":
if r == 'Male':
entry = 1
else:
entry = 0
elif h == "age":
# age
age = int(r)
if age <= 25:
entry = 0
elif age <=65:
entry = 1
else:
entry = 2
elif h == "race":
# race
if races.index(r) == 3:
entry = 1
else:
entry = 0
#entry = getHotEncoding(races.index(r), len(races))
elif h == "priors_count":
# priors count
priors = int(r)
if priors <= 0:
entry = 0
elif priors <=10:
entry = 1
elif priors <=20:
entry = 2
elif priors <=30:
entry = 3
elif priors <=40:
entry = 4
else:
entry = 5
elif h == "days_in_jail":
# months in jail
months = int(r)/12.0
if months <= 0:
entry = 0
elif months <=3:
entry = 1
elif months <=6:
entry = 2
elif months <=12:
entry = 3
elif months <=24:
entry = 4
elif months <=48:
entry = 5
elif months <=60:
entry = 6
else:
entry = 7
elif h == "c_charge_degree":
entry = fmo.index(r)
else:
entry = float(r)
domainArray[j].add(entry)
try:
temp.extend(entry)
except:
temp.append(entry)
nrecords.append(np.array(temp))
nrecords = np.array(nrecords)
for index in range(19):
temp = []
for record in nrecords:
temp.append(record[index])
domainArray = [np.array(list(uvs)) for uvs in domainArray]
simpleDomain = Domain(head, domainArray)
return simpleDomain, nrecords
def test_compas():
# just test that there are no errors for default loading...
cd = CompasDataset()
def LoadData(dataset_name,protected_attribute_name,raw=True):
optim_options=None
if dataset_name == "adult":
if raw:
dataset_original = AdultDataset()
if protected_attribute_name == "sex":
privileged_groups = [{'sex': 1}]
unprivileged_groups = [{'sex': 0}]
if not raw:
dataset_original = load_preproc_data_adult(['sex'])
optim_options = {
"distortion_fun": get_distortion_adult,
"epsilon": 0.05,
"clist": [0.99, 1.99, 2.99],
"dlist": [.1, 0.05, 0]
}
elif protected_attribute_name == "race":
privileged_groups = [{'race': 1}]
unprivileged_groups = [{'race': 0}]
if not raw:
dataset_original = load_preproc_data_adult(['race'])
optim_options = {
"distortion_fun": get_distortion_adult,
"epsilon": 0.05,
"clist": [0.99, 1.99, 2.99],
"dlist": [.1, 0.05, 0]
}
elif dataset_name == "german":
if raw:
dataset_original = GermanDataset()
if protected_attribute_name == "sex":
privileged_groups = [{'sex': 1}]
unprivileged_groups = [{'sex': 0}]
if not raw:
dataset_original = load_preproc_data_german(['sex'])
optim_options = {
"distortion_fun": get_distortion_german,
"epsilon": 0.05,
"clist": [0.99, 1.99, 2.99],
"dlist": [.1, 0.05, 0]
}
elif protected_attribute_name == "age":
privileged_groups = [{'age': 1}]
unprivileged_groups = [{'age': 0}]
if not raw:
dataset_original = load_preproc_data_german(['age'])
optim_options = {
"distortion_fun": get_distortion_german,
"epsilon": 0.05,
"clist": [0.99, 1.99, 2.99],
"dlist": [.1, 0.05, 0]
}
dataset_original.labels = 2 - dataset_original.labels
dataset_original.unfavorable_label = 0.
elif dataset_name == "compas":
if raw:
dataset_original = CompasDataset()
if protected_attribute_name == "sex":
privileged_groups = [{'sex': 0}]
unprivileged_groups = [{'sex': 1}]
if not raw:
dataset_original = load_preproc_data_compas(['sex'])
optim_options = {
"distortion_fun": get_distortion_compas,
"epsilon": 0.05,
"clist": [0.99, 1.99, 2.99],
"dlist": [.1, 0.05, 0]
}
elif protected_attribute_name == "race":
privileged_groups = [{'race': 1}]
unprivileged_groups = [{'race': 0}]
if not raw:
dataset_original = load_preproc_data_compas(['race'])
optim_options = {
"distortion_fun": get_distortion_compas,
"epsilon": 0.05,
"clist": [0.99, 1.99, 2.99],
"dlist": [.1, 0.05, 0]
}
protected_attribute_set={
'sex':[[{'sex': 1}],[{'sex': 0}]],
'age':[[{'age': 1}],[{'age': 0}]],
'race':[[{'race': 1}],[{'race': 0}]]
}
if optim_options==None:
print('No such dataset & group option:', dataset_name, protected_attribute_name)
exit()
return dataset_original,protected_attribute_set[protected_attribute_name][0],protected_attribute_set[protected_attribute_name][1],optim_options
def load_preproc_data_compas(protected_attributes=None):
def custom_preprocessing(df):
"""The custom pre-processing function is adapted from
https://github.com/fair-preprocessing/nips2017/blob/master/compas/code/Generate_Compas_Data.ipynb
"""
df = df[[
'age', 'c_charge_degree', 'race', 'age_cat', 'score_text', 'sex',
'priors_count', 'days_b_screening_arrest', 'decile_score',
'is_recid', 'two_year_recid', 'c_jail_in', 'c_jail_out'
]]
# Indices of data samples to keep
ix = df['days_b_screening_arrest'] <= 30
ix = (df['days_b_screening_arrest'] >= -30) & ix
ix = (df['is_recid'] != -1) & ix
ix = (df['c_charge_degree'] != "O") & ix
ix = (df['score_text'] != 'N/A') & ix
df = df.loc[ix, :]
df['length_of_stay'] = (
pd.to_datetime(df['c_jail_out']) -
pd.to_datetime(df['c_jail_in'])).apply(lambda x: x.days)
# Restrict races to African-American and Caucasian
dfcut = df.loc[~df['race'].isin(
['Native American', 'Hispanic', 'Asian', 'Other']), :]
# Restrict the features to use
dfcutQ = dfcut[[
'sex', 'race', 'age_cat', 'c_charge_degree', 'score_text',
'priors_count', 'is_recid', 'two_year_recid', 'length_of_stay'
]].copy()
# Quantize priors count between 0, 1-3, and >3
def quantizePrior(x):
if x <= 0:
return '0'
elif 1 <= x <= 3:
return '1 to 3'
else:
return 'More than 3'
# Quantize length of stay
def quantizeLOS(x):
if x <= 7:
return '<week'
if 8 < x <= 93:
return '<3months'
else:
return '>3 months'
# Quantize length of stay
def adjustAge(x):
if x == '25 - 45':
return '25 to 45'
else:
return x
# Quantize score_text to MediumHigh
def quantizeScore(x):
if (x == 'High') | (x == 'Medium'):
return 'MediumHigh'
else:
return x
def group_race(x):
if x == "Caucasian":
return 1.0
else:
return 0.0
dfcutQ['priors_count'] = dfcutQ['priors_count'].apply(
lambda x: quantizePrior(x))
dfcutQ['length_of_stay'] = dfcutQ['length_of_stay'].apply(
lambda x: quantizeLOS(x))
dfcutQ['score_text'] = dfcutQ['score_text'].apply(
lambda x: quantizeScore(x))
dfcutQ['age_cat'] = dfcutQ['age_cat'].apply(lambda x: adjustAge(x))
# Recode sex and race
dfcutQ['sex'] = dfcutQ['sex'].replace({'Female': 1.0, 'Male': 0.0})
dfcutQ['race'] = dfcutQ['race'].apply(lambda x: group_race(x))
features = [
'two_year_recid', 'sex', 'race', 'age_cat', 'priors_count',
'c_charge_degree'
]
# Pass vallue to df
df = dfcutQ[features]
return df
XD_features = ['age_cat', 'c_charge_degree', 'priors_count', 'sex', 'race']
D_features = ['sex', 'race'
] if protected_attributes is None else protected_attributes
Y_features = ['two_year_recid']
X_features = list(set(XD_features) - set(D_features))
categorical_features = ['age_cat', 'priors_count', 'c_charge_degree']
# privileged classes
all_privileged_classes = {"sex": [1.0], "race": [1.0]}
# protected attribute maps
all_protected_attribute_maps = {
"sex": {
0.0: 'Male',
1.0: 'Female'
},
"race": {
1.0: 'Caucasian',
0.0: 'Not Caucasian'
}
}
return CompasDataset(
label_name=Y_features[0],
favorable_classes=[0],
protected_attribute_names=D_features,
privileged_classes=[all_privileged_classes[x] for x in D_features],
instance_weights_name=None,
categorical_features=categorical_features,
features_to_keep=X_features + Y_features + D_features,
na_values=[],
metadata={
'label_maps': [{
1.0: 'Did recid.',
0.0: 'No recid.'
}],
'protected_attribute_maps':
[all_protected_attribute_maps[x] for x in D_features]
},
custom_preprocessing=custom_preprocessing)
from classifiers.utils import all_binaries
#%%
# Specify the dataset
from aif360.algorithms.preprocessing.optim_preproc_helpers.data_preproc_functions import load_preproc_data_compas
from aif360.datasets import CompasDataset
#%%
TAU = 0.9
TRAIN_SPLIT = 0.7
SENSITIVE_ATTRIBUTE = 'sex'
DOMAIN = [1, 3, 3, 2, 1]
NAME = 'compas_small_{}'.format(SENSITIVE_ATTRIBUTE)
dataset = CompasDataset()
raw_dataset_df = dataset.convert_to_dataframe()[0]
raw_dataset_df['age1'] = (raw_dataset_df['age'] <= 25).astype(int)
raw_dataset_df['age2'] = ((raw_dataset_df['age'] > 25) &
(raw_dataset_df['age'] <= 65)).astype(int)
raw_dataset_df['age3'] = (raw_dataset_df['age'] > 65).astype(int)
raw_dataset_df['pior1'] = (raw_dataset_df['prior_count'] <= 0).astype(int)
raw_dataset_df['pior2'] = ((raw_dataset_df['prior_count'] > 0) &
(raw_dataset_df['prior_count'] <= 10)).astype(int)
raw_dataset_df['pior3'] = ((raw_dataset_df['prior_count'] > 10) &
(raw_dataset_df['prior_count'] <= 20)).astype(int)
raw_dataset_df['pior4'] = ((raw_dataset_df['prior_count'] > 20) &
(raw_dataset_df['prior_count'] <= 30)).astype(int)
raw_dataset_df['pior5'] = ((raw_dataset_df['prior_count'] > 30) &
(raw_dataset_df['prior_count'] <= 40)).astype(int)
raw_dataset_df['pior6'] = (raw_dataset_df['prior_count'] > 40).astype(int)
def get_data(dataset, protected_attribute, seed=101):
def protected_attribute_error():
raise ValueError(
f'protected attribute {protected_attribute} is not available for dataset {dataset}'
)
if dataset == 'adult':
from aif360.datasets import AdultDataset
dataset_orig = AdultDataset()
if protected_attribute == 'sex':
privileged_groups = [{'sex': 1}]
unprivileged_groups = [{'sex': 0}]
elif protected_attribute == 'sex_or_race':
dataset_orig.feature_names += ['sex_or_race']
dataset_orig.features = np.hstack([
dataset_orig.features,
np.expand_dims(
np.logical_or(*dataset_orig.features[:, [2, 3]].T).astype(
np.float64), -1)
])
dataset_orig.protected_attributes = np.hstack([
dataset_orig.protected_attributes, dataset_orig.features[:,
[-1]]
])
dataset_orig.protected_attribute_names += ['sex_or_race']
dataset_orig.privileged_protected_attributes += [np.array([1.])]
dataset_orig.unprivileged_protected_attributes += [np.array([0.])]
privileged_groups = [{'sex_or_race': 1}]
unprivileged_groups = [{'sex_or_race': 0}]
elif protected_attribute == 'race':
privileged_groups = [{'race': 1}]
unprivileged_groups = [{'race': 0}]
else:
protected_attribute_error()
elif dataset == 'german':
from aif360.datasets import GermanDataset
dataset_orig = GermanDataset()
if protected_attribute == 'sex':
privileged_groups = [{'sex': 1}]
unprivileged_groups = [{'sex': 0}]
elif protected_attribute == 'age':
privileged_groups = [{'age': 1}]
unprivileged_groups = [{'age': 0}]
else:
protected_attribute_error()
elif dataset == 'compas':
from aif360.datasets import CompasDataset
dataset_orig = CompasDataset()
if protected_attribute == 'sex':
privileged_groups = [{'sex': 0}]
unprivileged_groups = [{'sex': 1}]
elif protected_attribute == 'sex_or_race':
dataset_orig.feature_names += ['sex_or_race']
dataset_orig.features = np.hstack([
dataset_orig.features,
np.expand_dims(
np.logical_or(*dataset_orig.features[:, [0, 2]].T).astype(
np.float64), -1)
])
dataset_orig.protected_attributes = np.hstack([
dataset_orig.protected_attributes, dataset_orig.features[:,
[-1]]
])
dataset_orig.protected_attribute_names += ['sex_or_race']
dataset_orig.privileged_protected_attributes += [np.array([1.])]
dataset_orig.unprivileged_protected_attributes += [np.array([0.])]
privileged_groups = [{'sex_or_race': 1}]
unprivileged_groups = [{'sex_or_race': 0}]
elif protected_attribute == 'race':
privileged_groups = [{'race': 1}]
unprivileged_groups = [{'race': 0}]
else:
protected_attribute_error()
elif dataset == 'bank':
from aif360.datasets import BankDataset
dataset_orig = BankDataset()
if protected_attribute == 'age':
privileged_groups = [{'age': 1}]
unprivileged_groups = [{'age': 0}]
else:
protected_attribute_error()
else:
raise ValueError(f'{dataset} is not an available dataset.')
dataset_orig_train, dataset_orig_vt = dataset_orig.split([0.6],
shuffle=True,
seed=seed)
dataset_orig_valid, dataset_orig_test = dataset_orig_vt.split([0.5],
shuffle=True,
seed=seed)
return dataset_orig_train, dataset_orig_valid, dataset_orig_test, privileged_groups, unprivileged_groups
print(dataset.privileged_protected_attributes)
print('Unprivileged attribute values:')
print(dataset.unprivileged_protected_attributes)
print()
binary_label_metric = BinaryLabelDatasetMetric(
dataset,
privileged_groups=privileged_groups,
unprivileged_groups=unprivileged_groups)
print(f'Statistical parity difference: '
f'{binary_label_metric.statistical_parity_difference()}')
print(f'Disparate impact: {binary_label_metric.disparate_impact()}')
# classification_metric = ClassificationMetric(
# dataset,
# privileged_groups=privileged_groups,
# unprivileged_groups=unprivileged_groups
# )
# print(f'Equal opportunity difference: '
# f'{classification_metric.equal_opportunity_difference()}')
if __name__ == '__main__':
generate_fairness_report(CompasDataset(),
privileged_groups=[{
'race': 1
}],
unprivileged_groups=[{
'race': 0
}])