The Research And Improvement Of Bayesian Network Structure Based On The Hybrid Models

The classification problem is an important research field in data mining, and this kind of problem is very common and ordinary. For a long time, the claasification methed based on Bayesian network becomes one of the best way to solve such problems, such as Spam Detection, Product Recommendation, Text Categorization and so on. Bayesian classifiers can deal with fragmentary data effectively, and it has favourable classification accuracy. Although na?ve Bayesian classifiers, with simple structure, ignores the dependencies among attributes, the classification accuracy is still beyond other mature algorithms in some cases. Although na?ve Bayesian classifier has been applied to many fields to slove the problem of classfification, and the classification accuracy is good, however as classification problems become more and more complex, the performance of this kind classifiers with the hypotheis that ignoring attributes dependencies is not very effective.To make up for the shortcoming of na?ve Bayesian classifiers, many algorithms has been put forward. The main methods can be divided into three types: Attribute Selection, Structure Extend and Partial Learning. In addition to the above three methods, there some other methods, such as data set instance weight, attribute weight, multi-model combination and so on. But so far, research achievement of multi-model combination method is rare. This article puts forward new classification algorithms based on analysis and itegration of some mature Bayesian classification algorithms, and the new algorithms improve the classification accuracy of Bayesian classifiers.In this paper, we put forword a new Bayesian network classifiers, H-AODE. The algorithm “H-AODE” is based on AODE classifier and HNB classifier, it learn a tree augmented na?ve Bayesian classifier for each attribute, then constructs a virtual node on every model, the virtual node combines influence of all attributes to the current attribute node. At last, we do average on these TAN model. In order to verify the H-AODE classification effect, we do a large number of experiments. We select more than 40 representative data sets from UCI database, and compare H-AODE with AODE, HNB, KDB, NB and other mature classification algorithms. Finally, we do comparison and analysis on the test results.