A Transfer Learning Algorithm Based-on Bigraph

Text is a common form of Internet data.With the rapid development of the Internet,sum of the redundant text data contributes to great burden of the producers,managers and consumers,who use them so frequently.In order to solve this problem,automatic text classification is proposed for the automatic classification management of Internet text.Nevertheless,due to the characteristic of immediacy of the Internet text as well as a sharp difference between the old labeling text and the newly generated text,they do not apply to the independent and identically distributed in the feature space any more,which means that the annotated training model for automatic data directly is unsuitable for the new generation of classification tasks on data.Based on this,a new idea of knowledge transfer is proposed,which makes it possible for different but similar domains or tasks to use the old knowledge to transfer knowledge.However,at present,there are still problems in the migration learning algorithm,such as poor interpretability and efficiency.In the background of above problems,this paper summarizes the key techniques used in automatic text categorization and transfer learning,and then proposes a new learning algorithm based on the two part graph.The main idea of such algorithm is that after feature extraction and feature selection for text data,it combined documents and the characteristics of combined source domain and target domain construction document,featuring two maps.Secondly,based on two map,compute transfer relationship between two characteristics in the field to transfer the arbitrary relationship between features as the bridge of knowledge transfer.Thirdly,the characteristics of space target in the field of document feature space are mapped to the source in the field of the source domain,and after that use the classic machine learning classifier training model to mark the text.Finally,use source domain model in the target domain document automatic text classification.The results show that the proposed algorithm can effectively solve the problem of interpretation and efficiency improvement in the process of transfer learning through the experiments of theparameters,the classifier,the comparison and the explanation.