On Methods Of Projection And Graph Construction For Unsupervised Domain Adaptation

In the age of big data,it is time-consuming and laborious to collect and label enough data in practical application.In this case,it is an important research direction in machine learning to use a large number of labelled data from related source domain to assist in the execution of target domain tasks.The purpose of domain adaptation is to resolve the domain divergence between source domain and target domain,and to achieve knowledge transfer across domains.This paper is based on unsupervised domain adaptation,mainly focusing on the aspects of projection and graph construction in domain adaptation,and proposes algorithms with better classification performance.The main research contents and achievements are as follows:(1)As the traditional domain adaptation methods are difficult to achieve the projection of source domain and target domain,and to obtain the ideal low-dimensional representations,we proposed Coupled Projectiones Embedding sparse Subspace Domain Adaptation(CPESDA).In embedding subspace,the entire feature space can be better represented by sparsely reconstructing the target domain data using a combination of source domain and target domain data rather than using only incompleted source domain data while inheriting the robustness advantage of sparse representation.The introducing of embedding subspace coupled projections can solve the heterogeneous problem between different domains that a single projection cannot handle.The projection matrixs of the source domain and the target domain can be jointly learned to find the optimal subspace,which can effectively improve the classification performance.Finally,to deal with highly non-linear problems that cannot be solved by linear technology,the algorithm is changed to non-linear by kernel method.The algorithm is tested on several standard datasets and compared with classical domain adaptation methods to verify the effectiveness of the algorithm.(2)Considering that most domain adaptation methods based on feature transfer minimize domain divergence by minimizing the Maximum Mean Discrepancy(MMD),MMD only considers domain distribution differences but ignores the internal structure of the data.From the perspective of composition,we propose two different unsupervised domain adaptive methods,which are Manifold based Subspace Learning Domain Adaptation(MSLDA)and Discriminant Embedding Domain Adaptation(DEDA).Guided by the consistency of manifolds,MSLDA can maintain the geometric structure of manifolds in subspaces and reduce local regional differences.On the other hand,considering the distribution differences between the target domain and the reconstructed target domain,the MMD needs to be minimized to eliminate the domain divergence.Finally,regularizing the low-rank constraint of the reconstructed matrix can more effectively explore the global structure of data in different domains.DEDA constructes intrinsic graph and penalty graph in the domain adaptation framework based on the idea of graph embedding,to maintain the relationship between samples in the original space.From the perspective of graph embedding,minimizing MMD is considered as graph in the graph embedding framework.It is verified that minimizing the distribution of MMD between source and target domains will reduce the compactness within and within classes,which may lead to overlapping points in different domains,and data cannot be guaranteed feature discriminative.To solve the above problems,DEDA modifies the intrinsic graph,adds a weight correction graph to make the modified graph keep the compactness within classes while minimizing the distribution.At the same time,a penalty graph is constructed to increase the separability between classes and further enhance the discriminability of features.We experimented with these two algorithms on an open benchmark dataset and compared them with several classical domain adaptive methods.The experimental results verify that the two algorithms can effectively improve the classification accuracy.