Similarity-based Transfer Learning Algorithm Of BP Neural Network

Traditional machine learning algorithms not only require a large amount of tagged training data,but also require training data and test data to satisfy the same distribution.However,in practical applications,training data with the same distribution and labeling as the test data is usually lacking.Therefore,how to use the differently distributed and labeled training data in other fields to assist the completion of the target task is an urgent problem to be solved.Transfer learning is an effective way to solve this problem.In most transfer learning studies,there is a certain degree of similarity between the source domain and the target domain.Using similarity learning in the transfer learning framework can improve the effectiveness of transfer.The BP neural network algorithm is a widely used machine learning algorithm.However,when the labeled training data is small,the generalization ability of the model is reduced.Aiming at the above problems,this paper studies similarity-based transfer learning algorithm of BP neural network.The main research work is as follows:(1)A similarity-based single source transfer learning algorithm of BP neural network(TLBP for short)is proposed.Firstly,the algorithm obtains the optimal weight parameters of the source domain and the optimal similarity between a single source domain and the target domain.Secondly,the optimal weight parameter information in a single source domain in combination with the similarity is migrated to the target domain.Finally,a BP neural network model for the target domain is constructed.The BP neural network algorithm and the TrAdaBoost algorithm are used as the comparison algorithm.All algorithms were empirically analyzed using the Letter-recognition data set,the Wine Quality data set,and the 20 Newsgroups data set.The experimental results show that under most experimental combinations,the TL-BP algorithm has obvious advantages in average classification accuracy and classification time.However,the problem of "negative transfer" has appeared in individual experimental combinations.(2)Aiming at the "negative transfer" problem in TL-BP algorithm,a similarity-based multi-source transfer learning algorithm of BP neural network(MTL-BP for short)is proposed.The algorithm adds the number of source domains based on the TL-BP algorithm,and simultaneously transfers the optimal weight parameter information of each source domain in combination with the corresponding optimal similarity to the target domain.To verify the validity of the algorithm,experiments were performed on the Letter-recognition data set and the 20 Newsgroups data set.The final experimental results show that compared with the MultiSourceTrAdaBoost algorithm,the algorithm has obvious advantages in classification time and average classification accuracy.Compared with TL-BP algorithm and BP algorithm,the algorithm has higher average classification accuracy under the premise of close classification time.It shows the effectiveness of the algorithm and effectively avoids the problem of "negative transfer ".