Research And Application Of Unrestricted Bayesian Classifier Based On KDB Model

In recent years,with the rapid development of the information age,a huge amount of data has been produced.People can't judge what data to play a role in decision-making when processing the data,how to find knowledge from a large amount of data and get useful information is a problem that the data mining technology needs to solve.The classification problem in data mining is one of the research directions that researchers pay more and more attention to.The study of classification will help to make key decisions based on potential information.Bayesian network is based on the theory of Bayesian and graph theory,which is regarded as one of the most promising research methods in the fields of knowledge discovery,artificial intelligence and data mining.Bayesian classification model is a branch of the Bayesian network,which can be divided into two types: restricted and unrestricted.Naive Bayesian is the most classical restricted Bayesian classification model and a basic framework of other restricted Bayesian classification models,which is simple,fast and efficient,but its conditional independence assumption is not true generally in practice.On the basis of Naive Bayesian,the KDB classifier can further relax the independence assumption among the attributes,and can construct the network structure with arbitrary k order structure complexity.Although KDB has demonstrated remarkable classification performance,however,the restricted network structure makes it impossible to represent the Markov blanket of class variable,which corresponds to the optimal classifier.And the test instances are not fully utilized,the final decision thus may be biased,leading to the classification precision decreasing.In view of the above problems,this paper proposes an unrestricted k-dependence classifier based on the analysis of Markov blanket of class variable and local learning.From the aspects of structure complexity,classification effectiveness and computational efficiency,the main research is as follows:1.UKDB can express the network structure with arbitrary k order structure complexity along the attribute dependence spectrum,and can output two kinds of sub-classifiers,a global framework which describes the causal relationships implicated in training set and a local framework which describes the causal relationships implicated in test instance.The local framework can be considered a complementary part of the global framework.2.The experimental results on the UCI machine learning repository of 50 datasets indicate that the comprehensive performance of UKDB in zero-one loss,bias and variance is superior to KDB,and requires only relatively small computational complexity.3.In addition,the application of UKDB in breast cancer prediction and diagnosis is of great significance.The experimental results on Wisconsin breast cancer dataset show that the classification error rate of UKDB compared to KDB,significantly decreased by 54.1%.Overall,compared with KDB,UKDB can achieve good trade-off between structure complexity and prediction performance.