Research On The Weighted One-Dependence Bayesian Forest Based On AODE Model

The third wave of artificial intelligence was set off by the competition between Alpha Go and Li Shishi.Internet companies have started the layout in the field of artificial intelligence.Machine learning which is the core technology of artificial intelligence has also got more attention.The essence of machine learning is to learn useful rules from large collections of data.Bayesian network which is the branch of machine learning is widely applied in real life because of its good performance in causal reasoning and uncertain knowledge representation,for example,spam filtering,anticheat detection,credit evaluation,medical diagnosis and so on.Naive Bayes model(NB),which assumes that all attributes are conditionally independent,is the basic and simplest Bayesian network structure.Numerous algorithms have been proposed to improve NB by weakening its conditional attribute independence assumption,such as: Tree Augmented Naive Bayes(TAN),Averaged One-Dependence Estimators(AODE),k-Dependence Bayesian Networks(KDB).NB and TAN are more suitable for small datasets.KDB performs well on large data sets,but the executing time is too long.In contrast,AODE's classification accuracy and algorithm complexity are better than other algorithms,which is mainly due to its ensemble learning.AODE overcomes the attribute independence assumption by averaging over all models in which all attributes depend upon the class and a single other attribute which is called super parent.However,considering each attribute as a super parent magnify the dependencies between super parent and other attributes.Besides,averaging all models ignores how super parent affects the classification result.In order to inherit the advantages of the AODE model and make attribute dependence closer to the true distribution from sample data.The paper presents two Bayesian models: AODF and WAODF which are based on AODE model.Firstly,selecting the attribute which has weak dependence with super parent node and assigning new parent node for the chosen attribute,according to some principles.Once a new parent node is determined,a new sub Bayesian network is built,then a new classifier is generated.Secondly,each sub model in AODF is divided into several categories with the structural differences.The WAODF is presented by assigning attributes different weight values with mutual information and conditional mutual information respectively,according to the differences within and between categories for each sub model in AODF.In order to verify the classification accuracy and stability,extensive experiments and comparisons on 40 UCI data sets are carried out.Results of experiments demonstrate that the new approach performs well in terms of both 0-1 loss and macroaverage.