Multi-Source Transfer Regression via Source-Target Pairwise Segment

Abstract Transfer learning addresses the problem of how to leverage acquired knowledge from a source domain to improve the learning efficiency and accuracy of the target domain that has insufficient labeled data. Instead of one source domain, multiple domains could be the source domains that are available for knowledge transfer in practice. However, there are large differences between the source and target domains, how to extract the useful knowledge from these different source domains remains a problem. To solve this problem, we propose a source-target pairwise segment method for multi-source transfer regression (STPS-MTR). The STPS-MTR method adaptively segments the different source domains and the target domain into different similar parts, and it extracts the most similar part in different source domains as the transfer knowledge. The STPS-MTR method can effectively extract the transfer knowledge from different source domains even when the source domain and the target domain have relatively low similarity, and it can avoid the negative influence between different source domains to ensure the transfer performance. Experimental results using synthetic datasets and real-world datasets demonstrate that the proposed method has better performance than existing methods, particularly when there are significant differences between different source domains and the target domain.