Research About The Selective Naive Bayesian Classification Based On Weighted Attributes

Bayesian Classification is one of the most important classification algorithms in data mining,The Naive Bayes is a simple kind of Bayesian Classification. It has been widely applied because ofthe advantages compared to other methods, such as simple, fast, stable and solid theory foundation.But this classifier requires attributes are mutually independent with given feature of classification,obviously, this assumption is rarely true in many real applications thus become the method’sdisadvantages. Many researchers have attempted to alleviate the independence assumption so as toimprove the performance of NB, Attribute Weighting and Attribute Selection are two goodmethods.In this paper, both of Attribute Weighting and Attribute Selection are studied to improve theperformance of NB. The main research works in the paper include the following aspects:(1)Improve NB by Attribute Weighting, introduce the basic principle of Weighted NB detailed,and analyze the influence of weight for the result in-depth, then introduce a method to determinethe weights of Attributes called Related Probability and construct a Weighted NB according to thismethod.(2)Improve NB by Attribute Selection, analyze two different Attribute Selection methodsin-depth, attributes correlation measure and wrappers. About the first method, introduce a specificalgorithm using Chi Square Statistics and construct a selective NB according to the algorithm;about the second method, study on the process of selecting attributes deeply and its aspects thatneed to pay attention to, and then construct a selective NB according to this method.(3) Propose the further improved models called WRNBC and WRSNBC that combine twomethods. WRNBC is a model that combines Attribute Weighting and attributes correlation measure,first get the optimal reduction subset of attributes by Chi Square Statistics, then construct WeightedNB in this subset; while WRSNBC combines Attribute Weighting and two different AttributeSelection methods, the first step is the same as WRNBC, the second step is to choose attributesfurther using wrappers, then construct Weighted NB in the final subset.