Research On Heterogeneous Domain Adaptation Algorithm

With the development of modern society,the application scenarios and practical problems encountered by artificial intelligence become more and more complex,and the cost of obtaining sufficiently annotated data sets becomes more and more expensive,which makes the application difficulty of traditional supervised machine learning algorithms become more difficult.As an important branch of transfer learning,domain adaptation is a machine learning method that has attracted much attention in recent years.When the target domain data set is not labeled enough,the classifier can be trained with similar but different labeled source domain data and the target domain can be classified.In addition,in real life,there is another more common and challenging situation,that is,the heterogeneous domain adaptation problem caused by the inconsistency of the feature space between the source domain and the target domain.In view of the heterogeneous domains adaptation mentioned above,the following aspects are mainly studied in this paper:(1)For the problem of heterogeneous domain adaptation classification,we consider to expand the difference between classes on the basis of the joint training of two domains,and put forward Domain Invariant and Class Discriminative Heterogeneous Domain Adaptation.Firstly,the mapping matrix is learned for the source domain and the target domain respectively,and the data of the two domains are mapped to the shared subspace.Joint optimization of the mapping matrix is carried out.Secondly,the maximum mean difference distance is used to adapt the distribution.Then,the repulsive force term is introduced to further expand the differences between classes,which is convenient for the training of mapping matrix and classifier.Finally,the structural consistency constraint and feature selection were carried out to complete the construction of the overall optimization target,and the label prediction of the target domain samples was carried out.(2)It is found that most of the existing algorithms ignore the application of target domain marker data.Therefore,we turn our research focus to a small number of labels in the target domain data and propose heterogeneous domain adaptation based on the target data distribution.Firstly,we use the maximum mean difference distance to distribute the source domain and target domain,and learn the mapping matrix.Then,with the help of the maximum mean difference distance,the labeled part and the unlabeled part are aligned and adapted.The mapping matrix is optimized by using the characteristics of the target domain,and the differences between the two domains are narrowed.Finally,the parameter complexity of the mapping matrix is controlled.The overall optimization objective is constructed,and the target domain is classified and predicted.This paper conducts several groups of comparative experiments on Office-31 visual transfer data set,Caltech-256 object recognition data set and Multilingual Reuters Collection data set respectively.We conduct extensive experiments,and the experimental results demonstrated that our algorithm is significantly superior to existing superior algorithms.