Research On Domain Adaptation Methods For Modeling Multi-source Heterogeneous Data

Domain adaptation,as an important branch of transfer learning,it aims to transfer the knowledge learned from labeled source domain to target domain with different distribution and lack of labels,so as to help the target domain complete the learning task.At present,most methods only focus on single-source domain,and assume that the feature spaces of source domain and target domain are homogeneous,so their effectiveness can only be ensured under specific scenarios.However,in practical,more than one source domain may be available,and the feature spaces of different domains may be heterogeneous.Therefore,how to complete the domain adaptation task in more practical scenarios is worth further study.The purpose of this thesis is to study the domain adaptation method for multisource heterogeneous data.Therefore,in a step-by-step way,the most basic singlesource homogeneous domain adaptation is firstly considered in this thesis,and then the heterogeneous scenario is further considered,and finally the multi-source scenario is considered.At present,the existing works still have some deficiencies:(1)the existing single-source heterogeneous domain adaptation methods only align the marginal distributions of source domain and target domain,but ignore aligning the conditional distributions;(2)the existing single-source heterogeneous domain adaptation methods only map heterogeneous samples in the source domain and the target domain to a domaininvariant common feature space,but ignore the domain-specific information;(3)the existing of multi-source heterogeneous domain adaptation methods treat each source domain equally,and do not distinguish the importance of different source domains.To solve the above problems,the domain adaptation methods are studied gradually in this thesis,and the main work of this thesis is as follows:(1)In order to solve the problem that existing methods fail to align the marginal distribution and conditional distribution of two domains simultaneously,a single-source homogeneous domain adaptation method based on joint distribution alignment is proposed in the thesis on the basis of the existing adversarial methods.This method mainly improves the discriminator to make it able to distinguish the domain and category of the feature extracted by the feature extractor,and trains it with the feature extractor adversarially to align the marginal distributions and the conditional distributions of source domain and target domain,so as to improve the performance of domain adaptation.The experimental results show that this method is effective in single-source homogeneous domain adaptation tasks.(2)In order to solve the problem that existing methods fail to extract domainspecific information,a single-source heterogeneous domain adaptation method based on domain-specific information extraction is proposed in the thesis to on the basis of the first work.This method mainly introduces a auto-encoder for both source domain and target domain,which is used to extract domain-specific features,so as to help improve the performance of the final model.The experimental results show that this method is effective in single-source heterogeneous domain adaptation tasks.(3)In order to solve the problem that existing methods fail to distinguish the importance of different source domains,a multi-source heterogeneous domain adaptation method based on adaptive domain weighting strategy is proposed in the thesis on the basis of the above two works.This method mainly designs an adaptive domain weighting strategy,which gives appropriate weights to each source domain according to the distribution discrepancy between different source domains and the target domain,so that the transferable knowledge from different source domains can be transferred to the target domain more effectively.The experimental results show that this method is effective in multi-source heterogeneous domain adaptation tasks.