Finding a set of candidate parents using dependency criterion for the K2 algorithm

One of the most effective structure-learning methods in a Bayesian network is the K2 algorithm. Because the performance of the K2 algorithm depends on the node ordering, more effective node ordering inference methods are needed. In this paper, we introduce a novel method for finding a set of candidate parents for each node as input for the K2 algorithm. Based on the fact that the candidate parents are identified by estimated Markov Blanket, we first estimate the Markov Blanket of a variable by using the L1-regularized Markov Blanket. We then determine the candidate parents of a variable through its Markov blanket by introducing a new scoring function based on the dependency criterion. Then the candidate parents are used as input for the K2 algorithm for learning Bayesian network structure. Experimental results over most of the datasets indicate that the proposed method avoids creating extra edges and significantly outperforms the previous methods.