Research On An Improved Instance-Based Transfer Learning Algorithm

With the rapid development of science and technology,how to obtain valuable information in large amounts of data has gradually become a hot topic in machine learning research.The proposed transfer learning breaks through two limitations of traditional machine learning:training data and test data must obey the same distribution rules,and there must be a sufficient number of training data.Transfer learning can use the previously acquired knowledge to solve new but similar problems faster and more efficiently.Compared with traditional machine learning methods,transfer learning uses a large amount of training data in the easily available source domain to assist training of a small number of training data in the target domain to perform prediction model training,thereby overcoming the disadvantages of traditional machine learning,and improves the generalization ability of the prediction model,and greatly expands the scope of problems that can be solved by machine learning.In order to improve the performance of existing transfer learning methods to better handle the problem of knowledge transfer in real life,this thesis focuses on the relevant algorithms in the field of instance-based transfer learning,and addresses the problems existing in the classic TrAdaBoost algorithm in this field including:the weight of the source domain is falling too fast;the weight of the source domain is too low after the iteration is deepened;the difference between the base classifiers is small and the classification result of the final integrated model is not ideal,and four improvements are made.The improvements include:(1)A feature selection algorithm DEFSS based on differential evolution is proposed.The algorithm can generate different feature subsets for each base classifier based on the characteristics of different iteration weight distributions for each iteration of the training data.Training base classifiers on different feature subsets can increase the degree of difference between base classifiers;(2)In order to further increase the degree of difference between base classifiers,the heterogeneous base classifier technology is introduced for TrAdaBoost,using a variety of classification algorithms trains the base classifier;(3)According to the classification error of the base classifier in the source domain,a weighted factor is introduced to the source domain to slow down the decline in the weight of the source domain;(4)Set new initial weights to prevent source domain weights from going too low after iterative deepening.In order to verify that our improvements to TrAdaBoost are effective,we have named the improved algorithms DEFSS-ATrA,DEFSS-HeteroTrA.The comparison experiments of DEFSS-ATrA,DEFSS-HeteroTrA,TrAdaBoost,DTrA and ATrA are carried out on standard datasets 20Newsgroups and Reuters-21578 to verify the feasibility and effectiveness of the proposed algorithm DEFSS-ATrA and DEFSS-HeteroTrA.Experimental results show that the improved algorithm DEFSS-HeteroTrA achieves better classification accuracy than DEFSS-ATrA,TrAdaBoost,DTrA,and ATrA on the verification data set,which indicates that the improvement of TrAdaBoost in this thesis is effective.