Research On Dependence Analysis –based Algorithms For Bayesian Network Structure Learning

Bayesian Network is a very important kind of probability graphical model, It uses visual figure structure describes the conditions of the independent relationship between random variables, It has unique advantages of uncertainty knowledge expression and reasoning.Bayesian network has been widely used in machine learning, medical diagnosis, economic forecasting, artificial intelligence, data mining, statistical inference and other fields.However, when a large number of variables is involved in the domain, constructing the model becomes very difficult only by the experts knowledge. Based on this, Bayesian network structure learning from date and inference have become important and difficult problems in the field.For the methods based on independence test, the Bayesian network is considered to the model of network of independent variables. The methods first calculate the mutual information between nodes and conditional independence test to find out in the data set D conditional independence relationship between each variable, and then find out the conditional independence of consistent network model. The complexity of the algorithm mainly lies in the mutual information among computing nodes and conditional independence test. As the number of nodes increases, independence test times will grow. Dimension reduction for the complex model and a high computational efficiency of algorithm are necessary.In this paper, Bayesian network structure learning is introduced in comprehensive review, the research situations to Bayesian network structure learning are summarized for all kinds of algorithms, and points out the advantages and disadvantages of those algorithms. Constructing Bayesian network structures from data is an NP-hard problem. Thus, we propose an approach to solve this problem, which is based on mutual information and PC algorithm. This algorithm first obains an initial undirected graph by using mutual information, and learns a PDAG using PC algorithm.Experimental results show that our method outperforms the PC algorithms under the same conditions. The algorithm decreases the running time and the order of CI tests drastically compared with the PC algorithm.