A Research On Domain Adaptation Based On Low Rank And Graph Embedding Techniques

Conventional machine learning algorithms can learn well-adaptive models under two major prerequisites:(1)sufficient labeled training data are available?(2)training and test-ing data have the same probability distribution.In most real-world applications,however,it is hard to satisfy above two prerequisites.In order to break through the bottleneck of traditional machine learning algorithms,more and more researchers pay close attention to domain adaptation algorithm that uses the labeled source data to learn unlabeled target data in which the source domain and target domain have different but related distributions.A lot of domain adaptation algorithms have proved that domain adaptation is a good way to solve the problem of knowledge transfer with different domains.Meanwhile,domain adaptation has been widely used in image recognition,text categorization and sentiment analysis problems.Most of the current domain adaptation methods are usually divided into two cator-gories,i.e.,feature extraction that learns a new feature representation to reduce the dis-tribution discrepancy,and instance-reweighting that adjusts data weights based on the correlation between source instances and target instances.Also,there are some joint optimization methods proposed combining the two methods,which further reduces the domain discrepancy.However,these domain adaptation methods only concentrate on re-ducing domain distances,but ignores the internal structure of the data and the correlation between data.Although some latent methods focused on the protection of data attributes,but ignore the domain discrepancy,which causes not very good results.Therefore,this paper will put forward a new method to jointly optimize data distribution alignment and data information protection.In order to protect a great deal of data information better,this thesis mainly focus on the low rank and graph embedding technologies,and make detailed researches on the two kinds of technologies.Low rank technology can protect the data properties,i.e.,the discriminative information of source data and the data properties of target data,and graph embedding can record the relationship between data.The approach based on low rank and graph embedding technologies can greatly enhance the degree of protection for different data information when domain discrepancy is reduced.Extensive experiments on sev-eral open benchmarks verify that the new approach show the superiority to the existing,well-established methods.In addition,this paper applies the proposed domain adapta-tion method to online data stream classification,and proposes a cross-domain data stream classification model.Experiments show that this model can solve the data distribution difference problem that the current data stream classification algorithms can't solve.