Research On Data Processing Based On Decision Tree In Cyber-physical Systems

The data in the Cyber-physical systems is generally derived by the physical device from the real world.And the data often has a realistic significance. When use decision tree to classify the data different error classification result in different cost loss. In this thesis we combine cost-sensitive and decision tree to solve this kind of problem. In addition, when the amount of data is too large, the construction of the final decision tree classifier will consume too much time, so we will use genetic programming to optimize it.In this thesis, we present a method callede ConSenDT. This method uses the ensemble classifier technology to built some based decision tree classifiers, and uses bayesian formula to calculate the minimum risk classification rules to modify the class label. Then we use the modified data to built the final decision tree classifier. When calculating the minimum risk, this thesis considered prediction accuracy of each decision tree classifier, the effects of it as a parameter is introduced into the calculation of minimum risk process. Experiments show that the improved algorithm can achieve the guarantee on the basis of the classification accuracy of classification after the expense of the loss, at the same time avoid the presence of the classification of costly mistakes.In this thesis, we present a method called ConSen DT-GP. This method use genetic programming to optimize the ConSenDT method.Through choose and crossover operation, leaving the higher fitness value of base classifiers as the final.When modify data operation class label, only when the final base classifier inconsistent category judgement when need to modify the class label according to the condition of minimum risk.Experiments show that when the data set is too big, the improved algorithm can effectively reduce the time consumed by a decision tree classifier building, at the same time to ensure that has less cost loss and higher classification accuracy.