Research Of Unsupervised Domain Adaptation Method Based Clustering And Balance

Transfer learning is one of the hot researches in machine learning.Its main idea is to leverage knowledge from the related source domain to help the learning of the target domain.As one of the important research directions of transfer learning,there are lots of researches on domain adaptation in recent years.The proposed domain adaptation methods mainly take the joint distribution of the source and target domains into consideration,which could match the feature between domains,and assume that both domains have the balanced distribution among categories.However,there are disadvantages as follows: 1.The clustering structure of the target domain is not considered in the proposed domain adaptation methods.2.If the source and target domains have imbalanced categories,i.e.the proportions of the classes in the source and target domains are different,distribution alignment may cause the incorrect classification of the target domain and even negative transfer.Therefore,the researches of this paper mainly include the following two aspects.First of all,the classic UDA methods do not consider the data cluster structure,which is one focus of traditional unsupervised learning.In this paper,we try to explore the cluster structure in UDA.Specifically,we propose a general transfer learning framework,named Clustering for Domain Adaptation(DAC),to explore the cluster structure of target data in the process of domain adaptation.DAC seeks a domain-invariant classifier by simultaneously reducing the distribution shifts between domains and exploring the cluster structure for target instances.The optimization of DAC adopts ADMM strategy,in which each iteration generates a closed-form solution.Empirical results demonstrate the effectiveness of DAC over several real datasets.Secondly,a Balanced Unsupervised Domain Adaptation method based on Classifier Discrepancy(CD?BDA)is proposed.In CD?BDA,the instance confidence is measured by the difference between the classifiers to assign appropriate weights to the instances,and the mini-max learning of classifier difference is used to reduce the negative impact of the imbalanced classes of domains,as well as the classification error of the boundary instances.The CD?BDA model is solved by using the Adam strategy.Experimental results show that the classification performance of images can be improved by CD?BDA.