Classification Method Based On Clustering Dynamic TrAdaBoost Transfer Learning

Nowadays,with the advancement of science and artificial intelligence,scientists are paying more and more attention to extracting and classifying the data from the amounts of data.However,we found that there are a lot of unlabeled data while little labeled data.And in the process of extracting and classifying data,marking data by person is very time consuming,which limits the behavior of manually marking data and the number of labeled data samples in the experiment.Based on this situation,transfer learning can solve these shortcomings.Transfer learning can use a large amount of unmarked data in related fields to help a small amount of marked data in the target domain.After the development in recent years,various transfer learning algorithms are proposed and has been widely used in machine learning,data mining and other fields.The traditional AdaBoost algorithm is iterated to get the weight of the feature.For the correct classification,we will give a lower weight value and give a higher weight value to the wrong classification.In recent years,based on the AdaBoost algorithm,people proposed a TrAdaBoost transfer learning algorithm to solve the problem of insufficient training.However,we found that the weight of TrAdaBoost transfer learning algorithm will appear the problem of polarization in the experiment,which will affect the progress of weight iteration.This paper presents a clustering dynamic TrAdaBoost migration learning algorithm to research the classification problem,the main research content is as follows:(1)For the problems of transfer learning in the process,thi s paper introduces the clustering dynamic TrAdaBoost transfer learning algorithm,including its basic idea,training process,specific process and experiment process.(2)For the weight problem of the TrAdaBoost algorithm,a method of adding dynamic factors in the transfer learning process is proposed.In the process of source domain weight updating,a dynamic factor is added to adjust dynamically according to the error in the target domain.(3)In the process of transfer learning,the source domain can help the learning of the target domain,but there are a large number of extremely dissimilar data in the source domain,which will greatly interfere with the accuracy of classification during the transfer process.Considering this situation,we propose to use the clustering algorithm to clean the data of the source domain so as to improve the accuracy of our classification.In this paper,we introduce the clustering dynamic TrAdaBoost transfer learning algorithm.We will experiment with it in the 20 Newsgroups data set,and prove the effectiveness and reliability of the proposed algorithm.Then it will be compared with the traditional AdaBoost algorithm and TrAdaBoost algorithm in the dataset to prove the accuracy and superiority of this algorithm.