def greedy_bayes(dataset, k=0, epsilon=0):
"""Construct a Bayesian Network (BN) using greedy algorithm.
Parameters
----------
dataset : DataFrame
Input dataset, which only contains categorical attributes.
k : int
Maximum degree of the constructed BN. If k=0, k is automatically calculated.
epsilon : float
Parameter of differential privacy.
"""
num_tuples, num_attributes = dataset.shape
if not k:
k = calculate_k(num_attributes, num_tuples)
attributes = set(dataset.columns)
N = []
V = set()
V.add(random.choice(dataset.columns))
print(
'================== Constructing Bayesian Network ==================')
for i in range(1, len(attributes)):
print('Looking for next attribute-parents pair.')
rest_attributes = attributes - V
parents_pair_list = []
mutual_info_list = []
for child in rest_attributes:
print(' Considering attribute {}'.format(child))
for parents in combinations(V, min(k, len(V))):
parents = list(parents)
parents_pair_list.append((child, parents))
# TODO consider to change the computation of MI by combined integers instead of strings.
mi = mutual_information(dataset[child], dataset[parents])
mutual_info_list.append(mi)
if epsilon:
sampling_distribution = exponential_mechanism(
dataset, mutual_info_list, epsilon)
idx = np.random.choice(list(range(len(mutual_info_list))),
p=sampling_distribution)
else:
idx = mutual_info_list.index(max(mutual_info_list))
N.append(parents_pair_list[idx])
V.add(parents_pair_list[idx][0])
print('========================= BN constructed =========================')
return N
def worker(paras):
child, V, num_parents, split, dataset = paras
parents_pair_list = []
mutual_info_list = []
if split + num_parents - 1 < len(V):
for other_parents in combinations(V[split + 1:], num_parents - 1):
parents = list(other_parents)
parents.append(V[split])
parents_pair_list.append((child, parents))
# TODO consider to change the computation of MI by combined integers instead of strings.
mi = mutual_information(dataset[child], dataset[parents])
mutual_info_list.append(mi)
return parents_pair_list, mutual_info_list