The Study Of The Method And Arithmetic Of Learning Bayesian Networks

Bayesian networks is one of the most efficient models in the fields of uncertainknowledge expression and inference .It has the following characteristics: the expressionform of graph model, partial and distributed study mechanism and directly perceivedinference ; applicable in expressing and analyzing uncertain and probability things andefficiently reasoning partial, inaccurate and uncertain knowledge or information. In thefield of graph model and data mining, the central issue and difficult point is how to learnBayesian networks and to accurately express valuable information in the data through theefficient methods and algorithm. The learning Bayesian networks mainly includes: structural and parameter learning,parameters can be fixed through networks structure and data sets, so structural learning isthe core of learning Bayesian network. The efficient structural learning is the basisconstructing the most efficient network structure. The paper illustrates the development and theoretical basis of Bayesian networks andthe application of classification and prediction, uncertain reasoning project and cause andeffect data mining, emphatically studying Bayesian networks learning theories;illustrating the main contents of Bayesian networks learning and its structural learningmechanism. This paper raises a new method of discrete Bayesian networks structurallearning based on prediction ability. The prediction ability is to predict accuracy rate, sothe same prediction ability is sufficient and essential to condition independence, thus theintroducing of prediction ability combines the existence and direction of arcs amongvariables. The method has the followings characteristics: learning efficiency and accuracyare high, the learning structure tends to be simplified, avoiding excessive combining ofdata, able to deal with incomplete data, unnecessary to order variables and with thefunctions of resisting noise data. The typical application of the Bayesian theory is theclassifying study. It does not classify an object into a certain class absolutely, butcalculating its probability and the class with the greatest probability is the one that theobject belongs to; in many cases all properties in the Bayesian classification functionpotentially, that is, one or several properties can not determine classification but all theproperties are involved in it .The properties of Bayesian network classification can bediscrete, continuous and mixed. This paper further studies the typical Bayesian classifiers, that is ,na?ve Bayesianclassifier ,TANC(tree augmented na?ve Bayesian)classifier and Bayesian networkclassifier based on MDL. Theoretically, Bayesian network classifiers and unionclassifiers have same classification ability, the core of realizing Bayesian networkclassifier is Bayesian network structural learning .The efficient Bayesian networklearning mechanism is the basis of constructing Bayesian network classifier. DiscreteBayesian structural learning method based on prediction ability forms Bayesian networkclassifier learning model. A contrastive experiment was conducted on UCI machinelearning data. The experiment shows that this method can be efficiently helpful forBayesian network classifier learning and high classification ability.