Structure Learning In Bayesian Networks And Construct MBNC Experimental Platform

With the development of IT, data-mining technology has been used widely in practice. Bayesian Network, as an effect knowledge-indicating way and probability-reasoning model, is a powerful decision analysis tool dealing with graph of uncertain attribute. Since there is a great deal of data in the real world, how to deal with these data and search for useful knowledge is of real significance.Bayesian Network is a DAG with probability noted, consisting of two parts--network topological structure and partial probability distribution. This article expounds concisely three important theories in Bayesian Network: indication, study and inference of Bayesian Network. In recent years data-mining based on BN has made good effects in some data model-constructing problems.Bayesian Network used in categorizing is called Bayesian Classifier, which is a special form of Bayesian Network for that both variable-choosing and state number have been decided with attribute nodes given and class node unknown. Bayesian classifier family has three common classifiers: NBC, TANC and BNC. The study of Bayesian Classifier includes structure study and parameter study and the latter is relatively easier.Establishing and structuring of Bayesian Network is a problem this article is trying to resolve. At present mainly JavaBayes soft package, Hugin Expert , Power Constructor, MSBNx ,Netica etc are in common use which have been all downloaded by the author for study and use .Hugin Expert and so on are limited editions only equal to an execution program , New arithmetic realization can not be accomplished on Power Constructor, MSBNx, Netica etc. for they can not provide source code . JavaBayes software based on Java language provides source code. WEKA system and JBNC system is explored by using Java language .But workload of programming in Java language is huge. Program debugging is relatively difficult and system expandability is baddish. Especially when it concerns program in the area of maths-physics statistics, original source code is of inferior readability, the operation on the defined multi-dimension is overloaded with details and of fallibility. Created structure can be shown only by using other software. The similar problem also exists in other downloaded Bayesian Network study software .NBT software package based on Matlab language is able to resolve the above-mentioned problem quite well. Matlab language is a superior language specializing in numerical value calculation and workload in it in programming is much less than that in Java language .It is convenient to debug a program and show the structure obtained from learn. The shortcoming of BNT is that it has no GUI. Eventually MBNC experimental platform has been explored in Matlab language by using basic functions provided by BNC.Several Bayesian Classifiers are realized simply on MBNC experimental platform: NBC, TANC based on MI criterion, BNC based on K2 algorithm and G2 algorithm. Accuracy evaluation of the established classifiers is superior to literature data .Experimental data indicates: the classifier structured with Matlab language is of excellent performance in the area of maths-physics statistics. MBNC experimental platform is of good expandability and is convenient for further and new research.Complete Bayesian Network's structure-learning is an NP-hard problem. Many researchers have suggested approximate algorithm and achieved a good effects.This article tries to improve Bayesian Network's structure-learning. NBC do not need learn structure and our job is to improve the structure-learning algorithm of TANC and BNC. New algorithm has been validated on MNBC experimental platform .The criterion to weigh superiority and inferiority is accuracy-evaluation of the structured Bayesian Classifier. Evaluation is made on standard data- set downloaded from UCI.The improvement on TANC structure-learning is by introducing a new criterion function-BIC abbreviated from Bayesian Information Criterion and the original MI abbreviated from mutual information criterion...