Bayesian networks are probabilistic graphical models that determine the cause-and-effect relationship between variables and include structural learning and parametric learning. The K2 algorithm is one of the best methods for learning structure in Bayesian networks for discrete variables. The performance of the K2 algorithm is strongly affected by the order of the input variables. Therefore, to achieve an accurate graph that describes the data, it is necessary to find an algorithm that provides a more accurate order of elements as 2K inputs. In this paper, first, the Markov cover of each variable is found using the incremental-decremental method, and then, based on the conditional frequencies and using the Dirichlet probability density function, the possible parents of each variable are selected from the Markov cover of each variable. The set of selected parents of each vertex is used as the input of the K2 algorithm, and the Bayesian network is obtained. The results of applying the proposed algorithm on several benchmark datasets and comparing it with other methods show that the proposed algorithm is much more efficient than other methods.