Algorithms for Large Bayesian Networks

BigBraveBN

BigBraveBN is an algorithm that is used for structure learning of large Bayesian networks. It restricts the search space by using Brave coefficient, that represents mutual occurrence of two variables in groups. These groups are formed for each variable using kNN algorithm that searches nearest neighbors for each variable. Mutual information score is used as metric for nearest neighbors algorithm.

BigBraveBN initialization and usage

To use BigBraveBN, just follow typical structure learning procedure with one difference: use BigBraveBN to generate white_list.

First, initialize BigBraveBN object and generate possible edges list:

space_restrictor = BigBraveBN()

space_restrictor.set_possible_edges_by_brave(
  df = data)

ps = space_restrictor.possible_edges

Then, preprocess the data:

encoder = preprocessing.LabelEncoder()
discretizer = preprocessing.KBinsDiscretizer(n_bins=5, encode='ordinal', strategy='uniform')

p = pp.Preprocessor([('encoder', encoder), ('discretizer', discretizer)])
discretized_data, est = p.apply(data)

info = p.info

Then perform structure learning as usual, but use ps as white_list:

bn = networks.ContinuousBN()

bn.add_nodes(descriptor=info)

params = {'white_list': ps}

bn.add_edges(discretized_data, scoring_function=('K2',K2), params=params)