The Research Of One-dependence Bayesian Model Based On Attribute Selection

During big data era,data information is one of the most valuable abstract objects which contains lots of valuable information which requires us to extract and thus makes data mining a major tool.Data mining,a popular discipline in our society nowadays,represents the process of discovering knowledge from stored data.It includes two directions: classification and clustering.This topic focuses on classification algorithm.Classification is a way to identify category based on a series of attributes and there are many classification methods.Recently Bayes network classification model receives increasingly more attention.Bayes classification model contains four classical classification models.Those are Naive Bayes(Na?ve Bayes,NB),Tree-Augmented Naive Bayes(Tree-Augmented Na?ve Bayes,TAN),Averaged One-Dependence Estimators(Averaged One-Dependence Estimators,AODE)and K-Dependence Bayesian(K-Dependence Bayesian,KDB)respectively.The core idea of Bayes classification lies in known prior distribution and then obtained posterior distribution from observed data,thus predicting classification.The proposed model belongs to One-Dependence Bayes model.The main idea is to sort attributes according to mutual information and non-class variable conditional mutual information to obtain an attribute sequence with its correlation from strong to weak and then build a classification model according to this attribute order.Local models are built in light of test cases during the process of testing.The model of GL_AS_TAN is a hybrid model which uses AS_TAN model on both global and local fronts.Global model is constructed for all attributes while local model is constructed according to individual situations.In this sense,hybrid models combine global models and local models to improve the accuracy of classification.K_TAN model also belongs to One-Dependence Bayes model with parameter K which represents the restricted number of selections when attribute parent nodes are selected for attributes.K_TAN_D,a combination of feedback system and K_TAN algorithm,builds a model adapted to each dataset by selecting suitable parameter K for each dataset through feedback system.This paper uses three metrics,0-1 loss,Bias and Variance,to compare the three models in this paper with other models and make relevant analysis.According to experiments,the proposed model is a clear improvement on classical One-Dependence Bayes model in classification accuracy.