Graph Embedding Based Visual Domain Adaptation

With the rapid growth of data scale,the use of machine learning to analyze and exploit data has achieved great success.However,in the fields of computer vision,pattern recognition and natural language processing,there are generally distribution discrepancy between different datasets,which will severely limit the practical application performance of traditional machine learning algorithms.Domain adaptation relaxes a basic assumption of traditional machine learning algorithms,that is,assuming that training data and test data are independent and identically distributed,which can effectively use knowledge in source domain to help promote learning tasks in the target domain.In this thesis,we have designed some novel learning strategies to address the problems and challenges of discriminative structure discovery,robust distribution adaptation and classifier transfer in domain adaptation.The main contributions are as follows:This thesis proposes an unsupervised domain adaptation algorithm based on coupled knowledge transfer.Firstly,to solve the problem of the high computational complexity of existing sample reweighting,a simple and effective he uristic sample reweighting strategy is designed to assign target predictions with different levels of confidence,so as to discover shared instances between domains.Secondly,to tackle the challenge of robust distribution adaptation,an improved weighted distribution adaptation is proposed to learn the shared features between domains.Finally,to make full use of manifold structure information within the domain,manifold regularization constructed by semantic information with different levels of confidence is proposed to promote the discriminant structure prediction and preservation.The proposed approach effectively incorporates these three terms into a general domain adaptation model,which promotes the transfer and preservation of label and structure knowledge from the source domain to the target domain,and from the high-dimensional input space to the low-dimensional feature space.For feature-based methods,the performance of basic classifiers may become the bottleneck of domain adaptation algorithms.To this end,this thesis proposes a new domain adaptation algorithm based on adaptive and discriminative transfer-classifier learning,which incorporates transfer-classifier learning and distribution adaptation into a general learning framework.In this algorithm,we are among the first to adopt the structural risk minimization principle on the target domain,combined with manifold regularization,which can make full use of geometric structure information of unlabeled data to enhance the process of classifier learning.Meanwhile,a weighted distribution adaptation based on the class probability matrix is designed,which can help to achieve a more robust and accurate alignment of joint probability distribution.Finally,we analyze the generalization error bound of the algorithm by the statistical learning theory,through which the correctness and effectiveness of the proposed algorithm are discussed theoretically.In this thesis,we have conducted comprehensive experiments on five benchmark datasets including Office-Caltech,COIL20,PIE,Office-Home and Image CLEF-DA to evaluate the performance of the proposed methods.We compare the proposed methods with 18 comparison methods on 124 cross-domain classification tasks,and the proposed methods achieve the highest classification accuracy of 106 tasks,as well as the best results in terms of the average performance on all datasets.Experimental results verify the superiority of the proposed methods to existing state-of-the-art domain adaptation methods.