A Bayesian network is a graphical model that represents a set of random variables and their causal relationship via a Directed Acyclic Graph (DAG). There are basically two methods used for learning Bayesian network: parameter-learning and structure-learning. One of the most effective structure-learning methods is K2 algorithm. Because the performance of the K2 algorithm depends on node ordering, more effective node ordering inference methods are needed. In this paper, based on the fact that the parent and child variables are identified by estimated Markov Blanket (MB), we first estimate the MB of a variable using Grow-Shrink algorithm, then determine the candidate parents of a variable by evaluating the conditional frequencies using Dirichlet probability density function. Then the candidate parents are used as input for the K2 algorithm. Experimental results for most of the datasets indicate that our proposed method significantly outperforms previous method.