Research On Domain Adaptation Methods For Manifold Structure Preservation

In classification learning,traditional single-domain classification models usually assume that training and testing data come from the same distribution.But in many realworld applications,changes in the external environment will lead to differences in the distribution of the collected data.Therefore,in scenarios different from the training data,current classification models usually cannot reach satisfactory effects.Domain adaptation does not require test and training samples to satisfy the IID assumption,which the question can be valid resolve.Domain adaptation establishes knowledge transfer from the source domain to the target domain by exploring domain-invariant knowledge structures,and the data distributions in these two domains are different.Most domain adaptation methods reduce distribution differences between domains by attempting to align cross-domain distributions,or perform manifold subspace learning to mine domain-invariant features and structure.However,these methods have the following major challenges:（1）when aligning cross-domain distributions,some domain adaptation methods do not evaluate the relative importance between the marginal distribution and conditional distribution.In addition,they do not consider class distinguishability,resulting in poor classification results.（2）During the mapping process,traditional graph-based regularization cannot strictly maintain the same nearest neighbor relationship with the original data,which brings about relatively weak representations.To this end,this thesis proposes a domain adaptation model of manifold regularization and adaptive joint discriminant distribution（MRAJDD）,which can effectively combine subspace learning methods and distribution alignment methods.Firstly,MRAJDD performs manifold feature learning to remove the threat of degenerate feature transformations.Then,to maximize domain transferability and class distinguishability,MRAJDD search a latent shared subspace in which the relative importance of the two probability distributions can be adaptively adjusted.And by introducing the repulsive force term,the distance between different label subdomains can be increased.Therefore,the discriminative ability of we proposed model is enhanced.Secondly,in order to maintain the inherent local geometric structure of the data and maintain the original proximity relationship of the data,MRAJDD proposes a manifold regularization constraint based on Euclidean distance calculation.Finally,a domain-invariant classifier is learned based on structural risk minimization,enabling the model to provide good generalization performance on the underlying distribution of the data.Through the research on existing domain adaptation algorithms,it is found that labels can effectively guide domain adaptation algorithms to perform feature transformation,and play an important role in learning highly discriminative projection matrices or classifiers.Based on this,this thesis further proposes a domain adaptation model of class distinction structure preservation（CDSP）.In order to better preserve the structural knowledge of the source domain samples in the same class and enhance the discrimination ability of the classifier,CDSP introduces the minimum intra class divergence into domain adaptation.Analyzing the local and label information of the samples,a novel manifold regularization is proposed by CDSP.While maintaining the local manifold structure of samples,this constraint maximizes the distance between source samples of different classes in the local area,so as to explore more discriminant features.Therefore,both models（MRAJDD and CDSP）are able to obtain discriminative and transferable feature representations to solve the problem of differences in data distribution.In order to verify the effectiveness of the two models proposed in this thesis.The two models proposed in this thesis are compared with the current popular benchmark methods in domain adaptation on four public datasets.The experiments show that the model proposed in this thesis has a significant performance improvement over the benchmark methods.