Research Of Transfer Learning Method Based On Partial Knowledge

Transfer learning,which applies the knowledge from related source domains to improve the learning of the target domains,is a hot research field in machine learning.At present,some methods consider selecting relevant instances to transfer or assume that the data categories are satisfied with balanced distributions,but it has the following disadvantages: 1)It ignores some related or unrelated knowledge for each instance in the source domain.For each source instance,possibly only part of the features is related for transfer.That is,when a source instance is selected by relativeness,its unrelated feature knowledge can also be introduced.At the same time,some related feature knowledge may be discarded when an unrelated source instance is dropped.2)If the category distribution is unbalanced in the source domain and the target domain,the judgment of the correlation between the source domain data and the target task will be incorrect.Negative transfer may occur when transfer knowledge from a less relevant source domain to target domain.As a result,the research content of this paper mainly includes the following two aspects:Firstly,we attempt to discover such partial related “instance-feature” knowledge in transfer,and propose a new transfer learning method with partial related “instance-feature” knowledge(Transfer Learning with Partial Related “Instance-Feature” Knowledge,PRIF).In PRIF,the partial “instancefeature” structure is first discovered by co-clustering,then the source instances are reconstructed by using this structural knowledge to improve the correlation between the source instances and the target ones.Experimental results show that PRIF can effectively enhance the learning performance of transfer learning.Secondly,a transfer learning method with local weight is proposed(Transfer Learning with Local Weight,TL_LW).In TL_LW,based on its ability of predicting accurately on each local region of the target domain,it assigns a proper weight to each local classifier,which is trained for each local region of the source domain,to alleviate the effect of imbalanced distribution.Experimental results show that the performance improvement can be obtained by the TL_LW.