The Design And Research Of Self-adaptive K-dependence Bayesian Classifier

Graphical models provide a principled approach to dealing with uncertainty andcomplexity. Graphical mode is very intuitive to show the relationship between attributes.Bayesian networks can clearly demonstrate causal relationship between attributes based onthe fundamental theorem of probability theory and Graphical models. Now combined withexpert knowledge and analysis of the data, Bayesian networks has become a hot research fieldof data mining. As one of the basic types of graphical models, restricted Bayesian classifiershave become an extremely popular tool for knowledge discovery.The scoring functions that were once proposed and widely used to evaluate therationality of Bayesian networks are inappropriate for Bayesian classifiers, where classvariable is particularly treated. Bayesian classifier includes many existing constructed models,which all have their own characteristics. Those existing models have shown differentperformance in particular data sets.This study established the mapping relationship between conditional probabilitydistribution and mutual information, and proposed to validate the rationality of Bayesianclassifiers from the perspective of information quantity implicated in a graphical structure.Combining the model construction method with characteristics of the Naive Bayes classifierNB, Tree augmented Bayesian classifier (TAN) and based on K-dependence Bayesianclassifier (KDB), To achieve global optimization and high dependence representation,proposing new learning algorithm, i.e. Self-adaptive K-dependence Bayesian Classifier(SKDB), applies greedy search in mutual information space to find the optimal order of theattributes and allows the construction of classifiers at arbitrary points (values of K) along theattribute dependence spectrum. Experimental results on45datasets in UCI machine learningdatabase repository proved the rationality of SKDB from the perspectives of zero-one lossbias and variance.