def process_my_own_algo_db(db_loc, epsilon, batch_id=None):
"""
Takes a synthetic database generated by the my_own_algo algorithm and
generates error statistics to be used in further analyses.
Copies the synthetic database into the data/test_output folder (in a
batch if specified) and calculates errors compared to the original
database. Also generates a summary file with the same format as
kendall analyses, however since there is no parameter `k`, leave
that field as 0.
"""
with open(params.attribute_file, 'r') as file:
attr_list = [attr.strip() for attr in file.readlines()]
# Create relevant directories and copy file
batch_dir = get_batch_directory(batch_id)
timestamp = datetime.now().strftime(TIMESTAMP_FMT)
epsilon_str = str(epsilon).replace('.', '-')
data_dir = f'{batch_dir}/my_own_algo_{timestamp}_{epsilon_str}'
makedirs(data_dir, exist_ok=True)
db_file = f'{data_dir}/synthetic_data.csv'
if not path.exists(db_file):
copy(db_loc, db_file)
# Load the synthetic data. This algorithm formats the database as a
# list of binary attributes with values of 1 if the row has the
# attribute and 0 otherwise.
synthetic_data = np.genfromtxt(db_file, dtype=int, delimiter=',')
original_db = Database()
original_db.load_from_file(params.input_data_file, params.attribute_file)
abs_errors = []
rel_errors = []
with open(f'{data_dir}/histogram_comparison.csv', 'w+') as file:
file.write('Attribute,Original count,Synthetic count,'
'Absolute error,Relative Error\n')
for index, attr in enumerate(attr_list):
original = original_db.one_way_histogram[attr]
synthetic = sum(synthetic_data[:, index])
abs_error = abs(synthetic - original)
rel_error = round(abs_error / original * 100, 2)
file.write(f'{attr},{original},{synthetic},{abs_error},'
f'{rel_error}%\n')
abs_errors.append(abs_error)
rel_errors.append(rel_error)
with open(f'{data_dir}/summary.txt', 'w+') as file:
file.write('epsilon,k,avg absolute error,avg relative error,'
'median absolute error,median relative error\n')
file.write(f'{epsilon},0,{np.mean(abs_errors)},'
f'{np.mean(rel_errors)},{np.median(abs_errors)},'
f'{np.median(rel_errors)}\n')
def test_get_marginal_histogram_shorter(self):
short_db = Database()
short_db.load_from_file('data/auxiliary/short_orig.csv',
'data/input/adult_columns.csv')
print(short_db.data)
print(short_db.numerical_attr_table)
marginal = short_db.get_marginal_histogram(0)
print(marginal)
dp_marginals = get_dp_marginals(short_db, 1)
print(dp_marginals[0])
def setup_files(self):
"""
Copies all data files required for the experiment into
the experiment folder.
Generates new Kendall data but reuses my_own_algo data as it takes
much longer to generate.
Copies files for all values of epsilon.
"""
super().setup_files()
for file in os.listdir(self.data_dir):
shutil.copy(f'{self.data_dir}/{file}', f'{self.exp_path}/{file}')
for file in os.listdir(self.attr_dir):
shutil.copy(f'{self.attr_dir}/{file}', f'{self.exp_path}/{file}')
original_db = Database()
original_db.load_from_file(self.original_data_src,
f'{self.exp_path}/adult_columns.csv')
for epsilon in self.epsilons:
eps1 = 8 / 9 * epsilon
eps2 = 1 / 9 * epsilon
kendall_data = kendall_algorithm(original_db, eps1, eps2)
kendall_db = Database(kendall_data, original_db.attr_table)
kendall_db.save_to_file(f'{self.exp_path}/'
f'kendall_synthetic_{epsilon}.csv')
def single_analysis(epsilon, k=8, batch_id=None, replace=False):
"""
Generates a single set of synthetic data and a one-way histogram
count of all attribute values.
epsilon: the total privacy parameter for the algorithm
k: the ratio of epsilon1 to epsilon2
batch_id: a str identifier to organise different analyses
replace: should replace analyses with same epsilon value
Results are saved into 'DPCopula/data/test_output/' with each
analysis with the same batch_id being saved in a batch_[batch_id]
folder.
"""
# Pick batch folder
batch_dir = get_batch_directory(batch_id)
makedirs(batch_dir, exist_ok=True)
if replace:
for data_dir in listdir(batch_dir):
if 'kendall' not in data_dir:
continue
eps = float(data_dir.split('_')[2].replace('-', '.'))
if round(eps, 6) == round(epsilon, 6):
rmtree(f'{batch_dir}/{data_dir}')
epsilon = round(epsilon, 6) # Ensure no floating point errors
epsilon1 = k / (k + 1) * epsilon
epsilon2 = 1 / (k + 1) * epsilon
epsilon_str = str(epsilon).replace('.', '-')
timestamp = datetime.now().strftime(TIMESTAMP_FMT)
output_dir = f'{batch_dir}/kendall_{timestamp}_{epsilon_str}'
makedirs(output_dir, exist_ok=True)
print(f'Running DPCopula-Kendall\tε = {epsilon}')
# Load data and create DP synthetic data
original_db = Database()
original_db.load_from_file(params.input_data_file, params.attribute_file)
synthetic_data = kendall_algorithm(original_db, epsilon1, epsilon2)
synthetic_db = Database(synthetic_data, original_db.attr_table)
synthetic_db.save_to_file(f'{output_dir}/synthetic_data.csv')
# Calculate error statistics
abs_errors, rel_errors = compare_databases(original_db, synthetic_db,
output_dir)
# Calculate summary statistics
with open(f'{output_dir}/summary.txt', 'w+') as file:
file.write('epsilon,k,avg absolute error,avg relative error,'
'median absolute error,median relative error\n')
file.write(f'{epsilon},{k},{np.mean(abs_errors)},'
f'{np.mean(rel_errors)},{np.median(abs_errors)},'
f'{np.median(rel_errors)}\n')
def generate_two_way_histogram(self, database_file):
"""
Generates a two-way histogram for a single synthetic database
specified by the name of the algorithm used and the value of
epsilon.
This operates identically to the function generating the one-way
histogram except the output file will have the format
'[attribute value 1],[attribute value 2],[count]'.
"""
database_type = database_file.split('_')[0]
if 'original' == database_type:
db = Database()
db.load_from_file(f'{self.exp_path}/{database_file}.csv',
f'{self.exp_path}/adult_columns.csv')
two_way_hist = db.generate_two_way_hist()
with open(f'{self.exp_path}/original_histogram.csv', 'w+') as file:
for pair, count in two_way_hist:
key1 = Database.avp_to_key(pair[0])
key2 = Database.avp_to_key(pair[1])
file.write(f'{key1},{key2},{count}\n')
elif 'kendall' == database_type:
epsilon = database_file.split('_')[2]
db = Database()
db.load_from_file(f'{self.exp_path}/{database_file}.csv',
f'{self.exp_path}/adult_columns.csv')
two_way_hist = db.generate_two_way_hist()
with open(f'{self.exp_path}/kendall_histogram_{epsilon}.csv',
'w+') as file:
for pair, count in two_way_hist:
key1 = Database.avp_to_key(pair[0])
key2 = Database.avp_to_key(pair[1])
file.write(f'{key1},{key2},{count}\n')
elif 'my_own_algo' == database_type:
epsilon = database_file.split('_')[2]
reordered_data = self.reorder_my_own_algo_data(database_file)
col_file = f'{self.exp_path}/adult_columns.csv'
attr_table = Database.get_attrs_from_file(col_file)
attr_val_pairs = [(attr, val) for attr in attr_table
for val in attr_table[attr]]
with open(col_file, 'r') as file:
attr_list = [attr.strip() for attr in file.readlines()]
with open(f'{self.exp_path}/my_own_algo_histogram_{epsilon}.csv',
'w+') as file:
for avp1 in attr_val_pairs:
for avp2 in attr_val_pairs:
if avp1[0] >= avp2[0]:
continue
key1 = Database.avp_to_key(avp1)
key2 = Database.avp_to_key(avp2)
col1 = attr_list.index(key1)
col2 = attr_list.index(key2)
count = sum(reordered_data[:, col1]
& reordered_data[:, col2])
file.write(f'{key1},{key2},{count}\n')
def generate_one_way_histogram(self, database_file):
"""
Generates a one-way histogram for a specific data file.
Parameters:
database_file = exact filename of the data file without the
extension
This assumes the database file has either of the following formats:
[algorithm]_synthetic_[epsilon].csv
original_data.csv
The file used will have the following format:
data/experiments/[folder_name]/[database_file].csv
The histogram will be stored in
[algorithm]_histogram_[epsilon].csv if the database is synthetic
or original_histogram.csv if it is the original database. Each
line has the format '[attribute value],[count]'.
If the database is my_own_algo synthetic, the columns will be
reordered as to generate the same histogram order as the
original and Kendall synthetic data produces.
"""
database_type = database_file.split('_')[0]
col_file = f'{self.exp_path}/adult_columns.csv'
with open(col_file, 'r') as file:
attr_list = [attr.strip() for attr in file.readlines()]
if database_type == 'original':
db = Database()
db.load_from_file(f'{self.exp_path}/{database_file}.csv',
f'{self.exp_path}/adult_columns.csv')
one_way_hist = db.generate_one_way_hist()
with open(f'{self.exp_path}/original_histogram.csv', 'w+') as file:
for attr in attr_list:
file.write(f'{attr},{one_way_hist[attr]}\n')
elif database_type == 'kendall':
epsilon = database_file.split('_')[2]
db = Database()
db.load_from_file(f'{self.exp_path}/{database_file}.csv',
f'{self.exp_path}/adult_columns.csv')
one_way_hist = db.generate_one_way_hist()
with open(f'{self.exp_path}/kendall_histogram_{epsilon}.csv',
'w+') as file:
for attr in attr_list:
file.write(f'{attr},{one_way_hist[attr]}\n')
elif database_type == 'my_own_algo':
epsilon = database_file.split('_')[2]
reordered_data = self.reorder_my_own_algo_data(database_file)
with open(f'{self.exp_path}/my_own_algo_histogram_{epsilon}.csv',
'w+') as file:
for index, attr in enumerate(attr_list):
file.write(f'{attr},{sum(reordered_data[:, index])}\n')
def setUpClass(cls):
cls.db = Database()
cls.db.load_from_file(params.input_data_file, params.attribute_file)