Research And Application Of Multi-classifier Selective Ensemble Metho

Ensemble learning methods enhance the ability of classification system by generating and combining multiple base classifiers.However,with the increase of the number of classifiers,the space and computing cost of the ensemble system will increase greatly.In view of this,some researchers proposed the selective ensemble method,which can not only reduce the storage and computing overhead,but also achieve better recognition rate and generalization ability than all the classifiers.The key of the selective ensemble algorithm is to evaluate the ability of the classifier on the verification set,so as to select the classifier members with the most ability to predict the unseen samples.How to correctly evaluate classifier capability and select appropriate classifier are urgent problems to be solved by selective ensemble method.This paper focuses on these two problems,and the main contents and work are summarized as follows:Firstly,a selective ensemble method based on clustering soft label and classifier sorting is proposed.In order to solve the problem that noise and redundant samples in the validation set samples have adverse effects on the capability evaluation of the classifier,the prediction results of the classifier on the validation set samples are expressed in oracle form.Then,a clustering soft label model is designed to perform sample space optimization and classifier clustering simultaneously,and the noise and redundant samples in the validation set are removed through sample optimization.Redundant learning is removed by classifier clustering.Finally,the best performance classifier is selected based on the prediction ability of classifier on the optimized verification set samples for fusion.Secondly,a selective ensemble method based on multi-label random walk is proposed.In this method,the problem of classifier selection is modeled as the problem of multi-label classification in order to flexibly and efficiently select classifiers.Firstly,the mapping between the training set sample and the classifier sequence is established,and the classifier sequence that can correctly identify the sample is regarded as the sample label.For a test sample,its nearest neighbor sample is found,and the multi-label random walk diagram is constructed according to the "many-to-many" relationship between the sample and the classifier sequence,and the random walk process is executed.According to the convergent probability vector,the corresponding classifier sequence is selected for ensemble.In this paper,experiments were conducted on FER2013,CK+,JAFFE,KDEF facial expression data sets and UCI data sets respectively.The experimental results show that the classification performance of the two models proposed is generally better than other comparative selective ensemble methods.At the same time,in order to embody the practical application of the classifier selective ensemble model,a face expression recognition system based on the front and back interaction is designed.The client mainly realizes the function of face collection and upload.The selective ensemble model is deployed to the server,the operation of facial expression recognition is performed,and then the recognition results are returned to the client.The final test results show that the expression recognition system has practicability,and can solve the computing and storage pressure of mobile devices to a certain extent.