Research On Bayesian Classifier Based On Volume Test

Data mining combines traditional method of data analysis with complicated algorithms which are meant to deal with large quantity of data, its aim is to mine pre-unknown but usful knowledge from data. Classification is a main branch in data mining. While Bayesian classification model as an important technology in classification field is based on statistical method. It has been widely used because of its simplicity and efficiency.This paper fistly introduces concepts of data mining and classification, including methods of construction and scoring a classifier. Secondly, it elaborates relevant concepts and theoretical basis of restricted bayesian classifier, summarizes different approaches of Bayesian learning, analyses their classical algorithms as examples. Then, comes to a conclusion that core problem in restricted Bayesian classifier learning is to find the relevant relationship among attributes. Consequently, deeply researches technologies of finding relevant relationship from data, especially compares Chi-Square test and Volume Test. Based on these, proposes a new algorithm named Tree Augmented Bayesian Classifier based on Volume Test. It constructs a Maximum Spanning Tree-like Bayesian network. But still has probability of improvement. Based on theoretical and experimental analysis, proproses another algorithm named Tree Augmented Bayesian Classifier based on combination. It aborbs advantages of Naive Bayes and statistics, in the mean time, combines idea of statistical hypothesis. Furthermore, after previous researchment, proposes a third algorithm named Bayesian Classifier using scoring based and conditional independence based method. It incarnates the advantages of two different learning approaches, and improves the flexsibility of Bayesian network. At last, as a key step in all three new algorithms of this paper, based on experimental analysis, proposes using the first SuperParent as root of Maximum Spanning Tree.Finally, programes the comparision algorithms and all new algorithms of this paper, as testifies by experiment, the three new algorithms show higher accuracy than Naive Bayes and TAN. Howerer, the third new algorithm shows even the same classification accuracy as SuperParent, but uses less time.