Research On Approximate Algorithm For Bayesian Network Classifier

Bayesian network has been extensively applied to various fields and it is also acommonly used effective classification method as a classifier. Bayesian networkwithout any structural constraint owns high complexity, which has limited itsapplication in multiple aspects. In order to improve efficiency of Bayesian networkclassifier facing big dataset with many attributes and acquire more practical Bayesiannetwork model, it’s necessary to analyze optimized model study method. Throughapproximate treatment for Bayesian network classifier algorithm, the calculated amountcan be effectively reduced and satisfactory classification accuracy rate will be obtained.Bayesian network classifier can be divided into generative Bayesian networkclassifier and discriminative Bayesian network classifier according to the difference ofscoring function when Bayesian network structure is studied. They have their ownstrengths and weaknesses: generative model supports simple model calculation butaccuracy rate of the result is low; discriminative model possesses high accuracy rate butits calculation process is complex. As for generative model, property of BIC scoringfunction is analyzed in this paper, so as to reduce search space and accelerate searchspeed of the algorithm when Bayesian network structure is studied. By combining thisproperty with K2search strategy, K2-aBIC classifier algorithm is proposed in this paper.In this way, the search space is further reduced and efficiency of the classifier is betterimproved when Bayesian network model structure is studied by this classifier.Experiment shows that K2-aBIC classifier can increase the efficiency on the premise ofguaranteeing classification accuracy rate.In terms of discriminative model, this paper analyzes a method of transformingdiscriminative scoring function of classifier into the decomposable generative scoringfunction through approximate learning. Then aCLL classifier obtained via this method isanalyzed. The analytical result shows that such classifier has low stability, and the ideaof improving it with ensemble thinking is proposed. Finally, Bagging-aCLL classifieralgorithm is gained. Experiment proves that classification performance ofBagging-aCLL classifier is strengthened to some degree.Therefore, approximate learning of Bayesian network classifier can make a goodbalance between its accuracy rate and execution speed. That is, Bayesian network classification model of high fitting degree with practical situation can be obtained, andmeanwhile time required to study such model is also reduced.