Study On Source Domain Selection Strategy Of Transfer Learning Based On Similarity

Transfer learning has become a hot topic in machine learning.Before conducting transfer learning,choosing the appropriate source domain for transfer learning in the target domain can avoid multiple transfer learning attempts,improve transfer learning efficiency and alleviate the phenomenon of negative transfer.However,most of the existing researches on transfer learning are based on the global constraint that the similarity between the source domain and the target domain is high.In view of the above problems,this paper conducts an in-depth study on how to select the appropriate source domain based on similarity for the transfer learning of the target domain.The main research work is as follows:(1)This paper expounds and analyzes the similarity measurement theory,and finds that the distance measurement method MMD can measure the distribution difference,and it is a non-parametric method,which is simple to calculate and conforms to the research goal of this paper,and can be used to measure the similarity between domains.(2)A method for measuring the similarity between two domains is proposed,which is called the domain similarity rank(MMD-SR).MMD-SR combines the advantage of the distance measurement method of maximum mean discrepancy(MMD).Furthermore,the variance factor is taken into account by combining MMD and distance variance.In addition,MMD-SR can directly show the rank of similarity degree of a candidate source domain in the set of all candidate source domains.Experiments were carried out in the classic transfer learning algorithms Tr Ada Boost,TCA and BDA,using the double-moon artificial data set,CaltechOffice data set and 20 News Groups data set.The experimental results show that MMD-SR is effective in measuring the distribution difference between domains,and the domain similarity rank MMD?SR is positively correlated with the accuracy of the transfer algorithm.(3)Based on the proposed method MMD-SR,MMD-SR is applied to the problem of source domain selection in the set of candidate source domains,and a source domain selection strategy MMD-SR?SDSS based on MMD-SR is proposed.The effectiveness and feasibility of MMD-SR?SDSS method are verified by experiments on text data set 20 News Groups using single source transfer learning algorithms Tr Ada Boost,TCA and BDA and multi-source transfer learning algorithm MS-Tr Ada Boost and SL-MSTL.