Research On Multi-Source Transfer Learning Boosting

The research of machine learning has been developed for more than 60 years,and the application of artificial intelligence can be seen in all aspects of life.Among them,transfer learning,as an important scientific research field,can link the source domain with the target domain,use the knowledge training model of the source domain,and apply it to the classification of the target domain.In recent years,the transfer learning algorithm based on domain adaptation has shown good results in the comparison of experiments.Compared with the traditional transfer algorithm,not only the accuracy has been greatly improved,but also the time efficiency of the algorithm has been improved.However,at present,most migration learning algorithms can only adapt to the training data of one source domain and one target domain,and cannot adapt to the samples of multiple different source domains.In addition,some transfer learning methods synchronously learn data from the source domain to the target domain model,and do not quantify the weight of learning knowledge according to the importance.In practice,the data divided according to one or some characteristics of a data set are often not uniformly distributed,and the importance of these different distributed data to the final model is also inconsistent,and the weight of knowledge transfer is therefore not equal.This paper conducts in-depth research on the transfer learning algorithm based on Boosting algorithm,and optimizes and improves the mainstream transfer learning Boosting algorithm from the perspective of multiple source domains and similarity difference transfer learning,and carries out relevant research work to improve the accuracy of multi-source domain transfer learning and expand the use scenarios of multi-source domain transfer learning.The research content of this paper mainly includes the following two aspects:(1)In order to realize the knowledge learning of multi-source domain transfer in a supervised scenario,a multi-source decision tree transfer boosting(Mtr Boost)method is proposed.The main idea is to calculate the KL distance to the target domain according to the source domain space of different distributions,and use the ratio of KL distance to calculate the learning weight proportion of the source domain samples of different distributions,so as to optimize the overall gradient function,so that the learning direction is the direction of the fastest gradient decline.Compared with other traditional transfer learning algorithms,the gradient descent algorithm can make the model converge faster,and ensure the learning speed while ensuring the transfer learning effect.(2)The above multi-source domain transfer learning accuracy is good,but it can only adapt to multi-source domain transfer learning in supervised scenarios.In order to realize multisource domain transfer learning in unsupervised scenarios,an unsupervised multi-source decision tree transfer boosting(UMtr Boost)method with optimized weights under multi-source domain distribution is proposed.This method determines the training data for labeling false labels according to the different distance between each source domain and the target.After labeling,different learning weights are given to different source domains Different samples are given different weights to improve the classification accuracy of the decision function.