def test_misses_cost_with_pgbs(self):
# Sensitive closed itemsets whose support needs to be reduced
sensitive_IS = {frozenset(['1', '2']), frozenset(['4'])}
# PGBS needs input in this format
sensitive_IL = pd.DataFrame(
{'itemset': [list(l) for l in sensitive_IS],
'threshold': [self.sigma_min, self.sigma_min]})
original_database = self.basket_sets.copy()
modified_database = self.basket_sets.copy()
# No return value, instead it modifies input database in place
pgbs(modified_database, sensitive_IL)
# Get all itemsets and supports in D (original_database)
sigma_model = 1 / len(original_database)
original_IS = fpgrowth(original_database, min_support=sigma_model, use_colnames=True, verbose=False)
# Get all itemsets and supports in D' (modified_database)
mofidied_F_IS = fpgrowth(modified_database, min_support=sigma_model, use_colnames=True, verbose=False)
# Find set of non-sensitive frequent itemsets in D
a = remove_sensitive_subsets(original_IS, self.sensitive_IS)
# Find set of non-sensitive frequent itemsets in D'
b = remove_sensitive_subsets(mofidied_F_IS[mofidied_F_IS["support"] >= self.sigma_min], self.sensitive_IS)
mc = misses_cost(a, b)
self.assertEqual(mc, 0.18181818181818182)
def test_artifactual_patterns_with_pgbs(self):
# PGBS needs input in this format
sensitive_IL = pd.DataFrame({
'itemset': [list(l) for l in self.sensitive_IS],
'threshold': [self.sigma_min, self.sigma_min]
})
original_database = self.basket_sets.copy()
modified_database = self.basket_sets.copy()
# No return value, instead it modifies input database in place
pgbs(modified_database, sensitive_IL)
# Get all itemsets and supports in D (original_database)
sigma_model = 1 / len(original_database)
original_IS = fpgrowth(original_database,
min_support=sigma_model,
use_colnames=True,
verbose=False)
# Get all itemsets and supports in D' (modified_database)
mofidied_F_IS = fpgrowth(modified_database,
min_support=sigma_model,
use_colnames=True,
verbose=False)
# All itemsets in original database
a = set(original_IS["itemsets"])
# All itemsets in sanitised database
b = set(mofidied_F_IS["itemsets"])
af = artifactual_patterns(a, b)
self.assertEqual(af, 0.0)
def test_information_loss_with_pgbs(self):
# Sensitive closed itemsets whose support needs to be reduced
sensitive_itemsets = pd.DataFrame({
'itemset': [['1', '2'], ['1', '2', '4'], ['4']],
'threshold': [self.sigma_min, self.sigma_min, self.sigma_min]
})
original_database = self.basket_sets.copy()
modified_database = self.basket_sets.copy()
# No return value, instead it modifies input database in place
print(sensitive_itemsets)
print(sensitive_itemsets.dtypes)
pgbs(modified_database, sensitive_itemsets)
# Give all itemsets and supports in D (original_database)
sigma_model = 1 / len(original_database)
a = fpgrowth(original_database,
min_support=sigma_model,
use_colnames=True,
verbose=False)
# Give all itemsets and supports in D' (modified_database)
b = fpgrowth(modified_database,
min_support=sigma_model,
use_colnames=True,
verbose=False)
il = information_loss(a, b)
self.assertEqual(0.5542, round(il, 4))
def test_pgbs(self):
basket_sets = im.import_dataset("chess").sample(100) # limit due to testing
original_database = basket_sets.copy()
modified_database = basket_sets.copy()
# We partition the Chess databases into 5 bins, then randomly select 2 itemsets from each bin,
# assign the minimum support threshold as the minimum support given in the support range
# This takes a long time, so will just use their values. Table 3: Support ranges for databases.
sigma_min = min([0.6001, 0.6136, 0.6308, 0.6555, 0.6974])
sigma_model = 0.5
original_IS = fpgrowth(original_database, min_support=sigma_model, use_colnames=True)
# Get 10 sensitive itemsets
sensitive_IS = original_IS.sample(10)
sensitive_IS_PGBS = pd.DataFrame({
'itemset': [list(IS) for IS in sensitive_IS["itemsets"]],
'threshold': [sigma_min for _ in sensitive_IS["support"]]})
pgbs(modified_database, sensitive_IS_PGBS)
# Give all itemsets and supports in D (original_database)
a = original_IS
# Give all itemsets and supports in D' (modified_database)
b = fpgrowth(modified_database, min_support=sigma_model, use_colnames=True)
il = information_loss(a, b)
self.assertEqual(0.5542, round(il, 4))
def test_hiding_failure_with_pgbs(self):
# Sensitive closed itemsets whose support needs to be reduced
sensitive_IS = {frozenset(['1', '2']), frozenset(['4'])}
# PGBS needs input in this format
sensitive_IL = pd.DataFrame({
'itemset': [list(l) for l in sensitive_IS],
'threshold': [self.sigma_min, self.sigma_min]
})
original_database = self.basket_sets.copy()
modified_database = self.basket_sets.copy()
# No return value, instead it modifies input database in place
pgbs(modified_database, sensitive_IL)
# Give all itemsets and supports in D (original_database)
sigma_model = 1 / len(original_database)
original_IS = fpgrowth(original_database,
min_support=sigma_model,
use_colnames=True,
verbose=False)
# Give all itemsets and supports in D' (modified_database)
mofidied_F_IS = fpgrowth(modified_database,
min_support=sigma_model,
use_colnames=True,
verbose=False)
# Find set of frequent itemsets in D
a = get_sensitive_subsets(
original_IS.loc[original_IS["support"] > self.sigma_min],
sensitive_IS)["itemsets"]
# Find set of frequent itemsets in D'
b = get_sensitive_subsets(
mofidied_F_IS.loc[mofidied_F_IS["support"] > self.sigma_min],
sensitive_IS)["itemsets"]
hf = hiding_failure(a, b)
self.assertEqual(0.0, hf)
def main(datasets):
#Create the base of a table
table_11 = pd.DataFrame(columns=['Model',
'Support threshold',
'Model threshold',
'Sensitive itemsets',
'Number of FI before sanitization',
'Information loss expected',
'Number of FI after sanitization',
'Number of FI containing an element of S after RPS',
'Hiding failure',
'Artifactual patterns',
'Misses cost',
'Side effects factor',
'Information loss',
'PGBS time'])
#Loop through datasets
for dataset in datasets:
sigma_model = datasets[dataset][0]
#Load dataset
data = im.import_dataset(dataset)
data = data.astype('bool') #This may be needed for some datasets
print("\n", dataset, "imported\n")
#Get frequent itemsets
freq_model = fpgrowth(data, min_support=sigma_model, use_colnames=True)
#Loop through support thresholds
for sigma_min in datasets[dataset][1:]:
print("\n", dataset, "FI:", sigma_min)
#Find original frequent itemsets at frequency sigma min
freq_original = freq_model.loc[freq_model["support"] >= sigma_min]
for k_freq in [10, 30, 50]:
print("-", dataset, ":", k_freq, "Sensitive itemsets")
#Copy the transactions so we can edit it directly
copied_data = data.copy()
#We pick sensitive itemsets here
sensitive_IS = get_top_k_sensitive_itemsets(freq_original, k_freq)
#Start timer for PGBS portion
total_time_start = time.time()
#Convert to pandas format for PGBS input
sensitive_IS_pandas = pd.DataFrame(data=[(sensitive_IS), np.full((len(sensitive_IS)), sigma_min)]).T
sensitive_IS_pandas.columns = ['itemset', 'threshold']
#Run PGBS
print("Running PGBS")
pgbs(copied_data,sensitive_IS_pandas)
print("PGBS run")
pgbs_time = time.time()
sensitive_IS = convert_to_sets(sensitive_IS)
print("FPGrowth")
#Reproduce frequent itemsets
freq_model_sanitized = fpgrowth(copied_data, min_support=sigma_model, use_colnames=True)
#Calculating metrics
#Variables needed
freq_sanitized = freq_model_sanitized.loc[freq_model_sanitized["support"] >= sigma_min]
#Sensitive subsets of frequent itemsets
freq_sanitized_sensitive = get_sensitive_subsets(freq_sanitized, sensitive_IS)
freq_original_sensitive = get_sensitive_subsets(freq_original, sensitive_IS)
#Non sensitive subset of frequent itemsets
freq_sanitized_nonsensitive = remove_sensitive_subsets(freq_sanitized, sensitive_IS)["itemsets"]
freq_original_nonsensitive = remove_sensitive_subsets(freq_original, sensitive_IS)["itemsets"]
#Calculation of metrics
freq_original_sensitive.to_csv("original.csv")
freq_sanitized_sensitive.to_csv("sanitized.csv")
print("len:", len(freq_original_sensitive["itemsets"]), len(freq_sanitized_sensitive["itemsets"]))
hiding_f = hiding_failure(freq_original_sensitive["itemsets"], freq_sanitized_sensitive["itemsets"])
artifactual_p = artifactual_patterns(set(freq_original["itemsets"]), set(freq_sanitized["itemsets"]))
misses_c = misses_cost(freq_original_nonsensitive.copy(), freq_sanitized_nonsensitive.copy())
side_effect_fac = side_effects_factor(set(freq_original["itemsets"]), set(freq_sanitized["itemsets"]), set(freq_original_sensitive["itemsets"]))
#Information loss between frequent itemsets in original and sanitized at sigma model
information_l = information_loss(freq_model.copy(), freq_model_sanitized)
#Expected information loss if all sensitive frequent itemsets had their support reduced to sigma min
expected_information_l = expected_information_loss(freq_model.copy(), freq_original_sensitive.copy(), sigma_min)
#Calculate the end time of this iteration
end_time = pgbs_time - total_time_start
#Threshold sanitized database by threshold_min to get frequent itemsets
print(f'- PGBS time: {end_time}')
#Plot support graphs
dual_support_graph_distribution(freq_model, freq_model_sanitized, sigma_model, dataset+"_PGBS_"+str(sigma_min)+"_"+str(k_freq))
#Find number of FI in sanitized database containing sensitive itemsets
num_FI_containing_S_RPS = count_FI_containing_S(freq_sanitized, sensitive_IS)
#Add to row of table
new_row = {'Model': dataset,
'Model threshold': sigma_model,
'Support threshold': sigma_min,
'Sensitive itemsets': k_freq,
'Number of FI before sanitization': len(freq_original),
'Information loss expected': expected_information_l,
'Number of FI after sanitization': len(freq_sanitized),
'Number of FI containing an element of S after RPS': num_FI_containing_S_RPS,
'Hiding failure': hiding_f,
'Artifactual patterns': artifactual_p,
'Misses cost': misses_c,
'Side effects factor': side_effect_fac,
'Information loss': information_l,
'PGBS time': end_time}
#Update after each one just so we are sure we are recording results
table_11 = table_11.append(new_row, ignore_index=True)
table_11.to_csv('table_pgbs.csv')
def main(datasets, experiment):
for dataset in datasets:
sigma_model = datasets[dataset][0]
sigma_min = datasets[dataset][1]
k_freq = 10
#Load dataset
data = im.import_dataset(dataset)
data = data.astype('bool') #This may be needed for some datasets
print("\n", dataset, "imported\n")
#Convert to closed itemsets
current_model, freq_model = get_closed_itemsets(data, sigma_model)
freq_original = freq_model.loc[freq_model["support"] >= sigma_min]
sensitive_IS = get_top_k_sensitive_itemsets(freq_original, k_freq)
if experiment == "MuRPS-range":
#Convert to pandas format for MRPS input
sensitive_IS_pandas = pd.DataFrame(
data=[(sensitive_IS),
np.array([
0.8, 0.79, 0.78, 0.77, 0.76, 0.75, 0.74, 0.73, 0.72,
0.71
]),
np.array([
0.795, 0.785, 0.775, 0.765, 0.755, 0.745, 0.735,
0.725, 0.715, 0.705
])]).T
elif experiment == "MuRPS-set":
#Convert to pandas format for MRPS input
thresholds = [
0.7975, 0.7875, 0.7775, 0.7675, 0.7575, 0.7475, 0.7375, 0.7275,
0.7175, 0.7075
]
sensitive_IS_pandas = pd.DataFrame(data=[(sensitive_IS),
np.array(thresholds),
np.array(thresholds)]).T
elif experiment == "SWA-set":
db = im.convert_to_transaction(data)
thresholds = [
0.7975, 0.7875, 0.7775, 0.7675, 0.7575, 0.7475, 0.7375, 0.7275,
0.7175, 0.7075
]
#Convert to pandas format for SWA input
sensitive_rules = get_disclosures(sensitive_IS, freq_model,
thresholds)
print(sensitive_rules)
#Run SWA
SWA(db, sensitive_rules, db.shape[0])
#Convert to frequent itemsets
sensitive_IS = convert_to_sets(sensitive_IS)
data = im.convert_to_matrix(db)
freq_model_sanitized = fpgrowth(data,
min_support=sigma_model,
use_colnames=True)
freq_sanitized = freq_model_sanitized.loc[
freq_model_sanitized["support"] >= sigma_min]
elif experiment == "PGBS-set":
thresholds = [
0.7975, 0.7875, 0.7775, 0.7675, 0.7575, 0.7475, 0.7375, 0.7275,
0.7175, 0.7075
]
sensitive_IS_pandas = pd.DataFrame(
data=[(sensitive_IS),
np.full((len(sensitive_IS)), thresholds)]).T
sensitive_IS_pandas.columns = ['itemset', 'threshold']
#Run PGBS
pgbs(data, sensitive_IS_pandas)
#Convert to frequent itemsets
sensitive_IS = convert_to_sets(sensitive_IS)
freq_model_sanitized = fpgrowth(data,
min_support=sigma_model,
use_colnames=True)
freq_sanitized = freq_model_sanitized.loc[
freq_model_sanitized["support"] >= sigma_min]
if experiment[0] == "M":
sensitive_IS_pandas.columns = [
'itemset', 'upper_threshold', 'lower_threshold'
]
print(sensitive_IS_pandas)
#Run RPS random threshold
sanitized_closed_IS = rps_two_thresholds(
model=current_model, sensitiveItemsets=sensitive_IS_pandas)
#Reproduce frequent itemsets
freq_model_sanitized = itemsets_from_closed_itemsets(
closed_itemsets=sanitized_closed_IS,
possible_itemsets=freq_model['itemsets'])
#Plot support graphs
dual_support_graph_distribution(
freq_model, freq_model_sanitized, sigma_model,
dataset + "_presentation_" + experiment + "_" + str(k_freq))
#Calculate and print information loss
information_l = information_loss(freq_model.copy(),
freq_model_sanitized)
print("Information loss:", information_l)