Research On The Weighted AODE Model Based On Association Rules

Entering the new century, with the advent of the big data era, data mining andclassification technology have been widely studied and applied.There are already a lot ofclassification methods having been applied to the field of data mining, which have played ahuge role. But it’s difficult for these techniques to give a causal relationship betweenattributes. Bayesian network is good at finding information through mutual dependence, andcan demonstrate a causal relationship by graphical network. It is an important means ofdealing with uncertain information, therefore it is very suitable for Bayesian networkclassification field.Currently, experts have made a lot of Bayesian classifier, the easiest restricted Bayesianclassifier is NB, but due to its limitation that assumes all properties are mutually independent,it is limited in practical applications. On the basis of NB model, scholars have put a lot ofclassification Bayesian classifier with high performance, such as TAN, KDB and AODE, etc.Although AODE has better classification performance, it takes the average of all classificationmodels which doesn’t consider that contributions to the classification of different models aredifferent. Besides, it does not consider the relevance between any two properties of the modelin addition to the catalog nodes and parent nodes.This paper presents two second-order dependent Bayesian classifier: AR-AODE andAR-WAODE with higher classification accuracy based on AODE. Since other attribute nodesare independent besides catalog nodes and parent nodes in AODE model, in order to alloweach attribute in Bayesian classification model to provide more information and raise theclassification accuracy, the paper uses the association rules algorithm Apriori to mine thecorrelation between different attributes. If two attribute nodes are correlated, then put themtogether, corresponding to the classification model, there will be an extended arc between twoattributes, then a new classifier AR-AODE is generated. Secondly, AODE is a classificationalgorithm which is the average of each classification model. Taking into account the reality ofclassification decisions, contributions to the classification accuracy of different models aredifferent, so this paper proposes a model named WAODE which values the different modelsthrough weighting to make the classification model be more reasonable. Finally, this papercombines the above two improved parts together, and presents the final new second-order dependent Bayesian classifierAR-WAODE.In order to verify the performance of the new classification algorithm, this paper uses50UCI data sets and1SEER data set for experiment testing.And it uses three indicators0-1loss,bias and variance to evaluation each classification performance. Experiments show that theproposed information enhanced second-order dependent Bayesian classifier has a morereasonable model structure and higher classification accuracy.