Research On Classifier Ensemble Based On Evidence Theory And Its Application

The classification algorithms based on different theoretical foundations can solve the classification tasks in a certain field,but there is no one method that can be well applied to all problems.Therefore,the research on classification algorithms is still a hot topic,and the integration of multiple classifiers is one of the essential directions.In addition,evidence theory has become an important part in the field of uncertain information fusion because of its obvious advantages in the expression,measurement and combination of uncertain information,and has been widely used in various kinds of research.It is also one of the combining strategies of classifier ensemble.This paper proposes a multi-classifier ensemble method based on the idea of D-S evidence theory.Specifically,firstly,complete the training of different individual classifiers.Secondly,the evaluation indicators of each classifier are regarded as the weights of evidence to quantify the value of each classifier in the ensemble.Then,these evaluation indicators are used to weight the outputs of individual classifiers through the weighting formula,and the weighted outputs are regarded as the improved evidence information.Finally,the evidence fusion is completed through the Dempster combination rule and the final classification results are obtained.For experiments,this paper chooses a prediction case of vehicle insurance and seven data sets in UCI.Five classification algorithms,including support vector machine,na?ve Bayes,gradient boosting machine,logistic regression,and adaptive boosting,are selected for ensemble,and two integration modes(three classifiers and five classifiers)are given.For each case,a comparative study among individual classifiers and the classification method proposed in this paper is presented,and the evaluation indicators,i.e.,accuracy rate,precision rate,recall rate and others,corresponding to each classification result are given too.These indicators are used to comprehensively evaluate the performance of the method proposed in this paper.In the case of predicting the possibility of purchasing vehicle insurance,the method proposed in this paper ensures that the value of AUC is close to the optimal,while significantly improving the precision rate and recall rate compared with other individual classifiers.Among the 7 cases of UCI,there are 5 cases reach or exceed the best value of evaluation indicators of individual classifications using the method proposed in this paper.The experimental results fully indicate that the method proposed in this paper,that is,the classifier ensemble method based on the evidence fusion weighted by the individual classifier evaluation indicators can improve the performance of classification and recognition,and also has good robustness and generalization.