Learning Bayesian Network Structure Based On Super Structure

With the rapid development of information technology,especially the emergence of big data,how to use the massive information reasonably to obtain useful information,has become a hot research topics of the data analysis and forecast.Bayesian networks can directly and vividly express the dependence relationship between data,has been applied to machine learning,medical diagnosis and information science,and many other areas.In the Bayesian network model,the nodes are used to represent the variables,and the directed edges are used to represent the causal dependence of variables.Aiming at the shortcoming of existing algorithms.A Bayesian network structure learning algorithm based on super structure is proposed in this paper.The main contents of the thesis are as follows.This paper firstly studies the framework of learning Bayesian Networks.When the data quantity is small or missing,the problem of the loss of the edge of the learning Bayesian network framework based on conditional independence test.This paper presents an efficient algorithm for learning Bayesian network framework.The algorithm first performs edge independent and first order independent inspection,and then use the necessary path condition identification which may be deleted edges during the conditional independence test,and finally get a super structure.When the framework is learning,the missing edge of the appearance is avoided.This algorithm is compared with the existing algorithms,experimental results show that the proposed algorithm has obvious improvement in time complexity and accuracy of the network framework.Secondly,the learning structure of Bayesian network is studied in the case of small sample.For small sample case,high order conditional independence test is not reliable lead to the low accuracy of the network structure.For this kind of problem,a new method of learning Bayesian network structure based on super structure is proposed in this paper,which is Opt01 HC algorithm.First of all,the use of super structure method to construct a undirected graph and can learn the appropriate redundancy edges,and well balanced number of missing edges and redundant edges,then the direction of the edge of the graph is obtained by the search algorithm,the structure of Bayesian network is obtained in finally.Simulation experiments are carried out on the standard data set ASIA,ALARM and INSURANCE.The results show that the learning efficiency and accuracy of the proposed algorithm are better than the MMHC algorithm and the accuracy of Opt01 HC algorithm is significantly increased with the increase of learning data.