Research Of Transfer Learning Algorithm Based On Semi-supervised Discriminant Analysis

The essence of traditional machine learning is advanced statistical applications,provided that the training data and test data require the same probability distribution.This premise limits the development of machine learning until transfer learning is proposed.The idea of transfer learning is to apply the learned knowledge to related fields to help accomplish the target task without requiring the data to be distributed.As human society moves into the era of artificial intelligence,the transfer learning that is closer to the learning of human beings can deal with the complex data in reality more effectively,which is of great research value.In this paper,we mainly study the method of transfer learning.Based on the theory of discriminant analysis,two different transfer learning algorithms are proposed based on the current research results.First of all,the traditional instance-based transfer learning methods are difficult to estimate distribution parameters,and the generalization effect is poor.Aiming at this problem,a transfer learning algorithm based on regularized discriminant analysis is proposed.Gaussian kernel and regular item are introduced into the linear discriminant analysis to be semi-supervised Gaussian kernel discriminant analysis method,which is more suitable for dealing with real data without considering the conditional probability density distribution.Constructing the discriminant space based on this method of discriminant analysis,reusable data in the source domain can be screened iteratively,which effectively avoids the estimation of distribution parameters.In order to avoid the occurrence of overfitting phenomenon,on the one hand,pseudo-mark data is constructed to assist screening so that the category information of mark data and the distribution information of unlabeled data can be fully utilized.On the other hand,defining distance metrics and indicating matrices during the screening process can effectively select the source data closest to the target domain data.To verify the effectiveness of the algorithm,experiments are performed using datasets such as 20 Newsgroups.The results show that the accuracy of transferring and the generalization of the learning model are effectively improved.Secondly,most feature-based transfer learning methods only focus on the inter-domain common features and neglect the unique features of each domain.Aiming at this problem,a transfer learning algorithm based on sparse local discriminant analysis is proposed.Combined with sparse local preserving projection,linear discriminant analysis is improved to a semi-supervised feature extraction method,which can make full use of local structure information and global supervision information of sample data.Using this method to establish the subspace of the source domain and the target domain can effectively extract the best sharing features among domains while maintaining the unique characteristics of each domain.To avoid excessive computational complexity,the method of aligning the subspace is adopted to reduce the difference between the domains so that the knowledge transfer can be realized.To prove the effectiveness of the algorithm,experiments are performed using datasets such as COIL20.The results show better performance of transferring.