Active Selection Transfer Learning Algorithm

• Published in: Neural processing letters Vol. 55; no. 7; pp. 10093 - 10116
• Main Authors: Wu, Weifei, Zhang, Yanhui, Xing, Fuyijin
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.12.2023; Springer Nature B.V
• Subjects: Accuracy; Active learning; Algorithms; Artificial Intelligence; Classification; Classifiers; Complex Systems; Computational Intelligence; Computer Science; Datasets; Fault diagnosis; Knowledge; Labeling; Labels; Machine learning; Probability distribution; Support vector machines; Support vector machine; Transfer learning; Active learning; Classification
• ISSN: 1370-4621; 1573-773X
• DOI: 10.1007/s11063-023-11240-1

Transfer learning has the ability to utilize the knowledge of the source domain with enough available and labeled data to help build a learning model for the target domain with scarce or unlabeled data, so it has gradually attracted the attention of researchers. This paper proposes an active transfer learning algorithm ACTrAda-TLSVM, the algorithm is based on support vector machine for the cross-domain data classification scenario where the source domain contains a small number of labeled samples and samples of the target domain are scarce enough to train a reliable classifier. It combines the advantages of active learning and enhanced learning to achieve the purpose of improving performance. The main components of ACTrAda-TLSVM are as follows: (1) utilizing maximum mean discrepancy to calculate the vector information of source domain sample weight and active learning to label the samples in the source domain with a large amount of information and similar to the target domain for purpose of obtaining high-quality label training samples; (2) on this basis, the knowledge in the source domain and data in the target domain are used to build the transfer learning classifier TLSVM, and (3) finally get TLSVM and TrAdaBoost together to construct ACTrAda-TLSVM classifier. Compared with the best benchmark algorithm, the experimental results show that the proposed algorithm improves by 1.74% and 1.21% on text datasets ISCX2012 and 20-Newsgroups, respectively; by 2.26% and 4.14% on image datasets COIL20 and Yale respectively. In addition, although the algorithm efficiency is considerable, there is still space for further improvement.