Research Of Classification Problem Based On Bart Algorithm

With the continuously improving of information technology,especially the rapiddevelopment of Internet technology,the size of information has a trend of explosive growth,soit become a common needs for almost all areas to mining valuable information from thosedata.As a most common problem in pattern recognition and machine leaning classificationproblem has always been drawn the resercher’s attention.At2008,By Hugh A. Chipmanproposed the baysian additive regression tree model,and expanded into classification problemin2010,it is a ensemble learning method that based on baysian tree,and has advantages ofgreat generalization ability and probability output.Due to the great efficiency in solvingtwo-class problem,we do some research about expanding this method to solve mul-classproblem, The main contents of the dissertation are as follows:Atfirst,this paper introduce three binarization strategies for solving multi-classificationproblem,and shows that it is possible to combine these strategies with BART classificationmodeland then propose a improved OAOBART algorithm(MOAOBART) to enhanceclassification ability with knowing the misclassified instance’s real class’s score rank has avery large percentage in second place, this method use membership to improve theclassification ability for the sample located into the unclassifiable regions. The test results ofthe algorithms to several UCI datasets show that,contrast to the originalalgorithm,MOAOBART algorithm improves a lot in classification accuracy of theunclassifiable regions and total classification accuracy.Because of the the rapid growth of the number of classifier with growth of the number ofclass,we propose a OAOBART algorithm based one time datadivision(ODMOAOBART).This method partion data into two parts with Approximately thesame number of class,and then for each part use MOAOBART to trackle,this method candecrease the number of classifiers to some extend,and also shorten the training and testingtime. The test results of the algorithms to ten UCI datasets show that, with losing littleaccuracy,the proposed algorithm has a less training and testing time.