Ensembling Bayesian network structure learning on limited data

In recent years, Bagging method has been applied to learn Bayesian networks (BNs), especially on limited datasets. However, the BNs learned using Bagging method from limited datasets can be biased towards complex models. We present an efficient approach to produce more accurate BNs from limited datasets. Based on the Markov condition of BN learning, we proposed a novel sampling method, called Root Nodes based Sampling (RNS), and a BNs fusion method. The experimental results reveal that our ensemble method can achieve more accurate results in terms of accuracy on limited datasets.