Research On Improving Naive Bayes Classification Model

Classification is an important task of data mining. The purpose of classification is to construct a classification function or classification model, which can map the unclassified sample in the database to a given class. Classification can be used to extract a model which describes important data or predicts the trend of data. Naive Bayes classification model is one of the research hotspots in current classification algorithms, and compared with other methods, Naive Bayes classification model owns features of simple structure, high classification accuracy and high speed, etc. Training set is used in Naive Bayes classification model to build a classification model, and if there are noise samples in the training set, the performance of the classification will be reduced. Taken optimizing the training set as research content, improved Naive Bayes classification model based on validity of single attribute and combined validity of double attributes are proposed. The noise samples in the training set are eliminated by validity of single attribute and validity of double attributes to achieve the goal of optimizing training set and improving classification accuracy.The main jobs are as follows:1. The basic theory of Bayes classification and the Naive Bayes classification model are introduced.2. Several common improved Naive Bayes classification model are analyzed: Semi Naive Bayes Classifiers (SNBC), Bayes Belief Network (BBN) and Tree Augments Naive Bayes (TAN).3. Based on Bayes theory, the noises of the training examples are eliminated by validity of single attribute to optimize the training set before it is used to build classifiers.4. Under the premise of Naive Bayes classification model based on validity of single attribute, an improved model combined validity of double attributes are proposed in order to discover and delete more noise samples.Experiment results based on mass data show that the proposed method in the dissertation is feasible, and they can effectively improve the classification accuracy.