The Research On Structural Learning Of Causal Bayesian Networks

Causal bayesian network is also called the belief network, It is the uncertainknowledge expression and reasoning model based on the graph theory ofprobability theory. It has important applications in data mining, pattern recognition,data compression, genetic information control, image processing, industrialmanufacturing. causal beyesian network is the basis of Parameter learning,Reasoning and classification, It is very important to study the method of learnstructure. So far, the structure learning of causal bayesian network is a NPproblem.The main works of this thesis are as follows:(1)This thesis makes a survey about current situations on structure learning,parameter learning, intervention learning and sensitivity analysis.(2) Mountain climbing algorithm and GS algorithm are difficult to get anapproximate global optimal causal bayesian network structure, and search moretimes. a greed search structure learning algorithm Based on2test (CIGSalgorithm) is presented. Initialized network was treated before mountain climbingalgorithms and GS algorithm learn Bayesian network. Firstly, we could study to geta undirected graph by2test; Then, use the conditionally relative averageentropy to decide the direction of undirected graph and make the direction ofnetwork learned more accurate. Finally, GS algorithm is used to learn Bayesiannetwork. The results of Experiments show that the algorithm can get a betterapproximate global optimal structure, reduce the search times and more efficient intime performance than mountain climbing algorithm and GS algorithm.(3) Algorithms based on independence test, such as PC algorithm, TPDAalgorithm are difficult to run in the large data sets of more nodes due to computingperformance. the concept of d-separation tree is introduced, It is usd to decomposea big bayesian network in to small networks, then these small networks are learnedby independence test method.finally we integrate these small networks.based onthis method, A improved structure learning algorithm is put forward based on thed-separation decomposition. this structure learning method of causal bayesiannetwork combine the advantages of search score algorithms and independence testalgorithms, it is more effective to learn the structure of causal bayesian network. (4) In this thesis we do some research on active learning areas, analyes thecausal power of Bayesian network and three models of intervention. finally, themutual information, symmetrical entropy, non-symmetric entropy are used tochoose nodes in structure learning. The effectiveness of the several methods arealso validated. At the same time, we put forward a node selection method to learnthe structure by combining mutual information and non-symmetric entropy. Theexperiment result show that in the small data sets this method greatly improve theeffect of active learning in the causal bayesian network.