Classification Methods Of Transfer Learning Based On Sparse Representation And Process Learning

Domain adaptation aims to help the learning of the target domain tasks by using the rich supervisory information of the source domain.Different from the traditional machine learning hypothesis,the two domains in domain adaptive learning,the source domain and the target domain,respectively follow similar but not exactly the same distributions.Domain adaptation is generally considered to have three learning methods.They are:(1)sample adaptive method,(2)feature adaptive method,(3)model adaptive method.In this paper,the first and second methods are explored in this paper,and the improved scheme to improve the classification effect of classification model is studied.The research scheme mainly includes the following two contents.Firstly,for the feature adaptive method,it is particularly important to retain the data characteristics before the mapping for the performance improvement in the common method of using the mapping subspace method to solve the distribution difference between the source domain and the target domain.Different from the common manifold structure characteristics,in this paper,the sparse structure characteristic which is more suitable for image classification is preserved in the construction of the model,and a domain adaptation method(DASSP)is proposed,which can be used in image classification with sparse structure preserving.Different from the manifold regularization which is reserved for the local structure information,this paper reconstructs the image samples by sparse representation and preserves the global relationship between the images.Experiments show the effectiveness of this method.Secondly,how to select and transfer the relevant knowledge is the key to learning in the sample adaptive method.A new domain adaptive model of process domain(PDA)is proposed in this paper.The model considers to transfer source knowledge based on the associated migration order.Specifically,PDA iteratively transfers the source instance from the most relevant source instance to the least relevant source instance until all relevant source instances are adopted.This is an iterative learning process in which the source instance used in each iteration is determined by a gradually increasing weight so that more source instances will be introduced into subsequent iterations.In addition,the reverse classification method is used to set the termination of the iteration.The experiments on the real data set show that PDA is competitive compared with the existing technologys.