Learning from Multiple Sources via Multiple Domain Relationship

Zhen LIU  Junan YANG  Hui LIU  Jian LIU  

Publication
IEICE TRANSACTIONS on Information and Systems   Vol.E99-D   No.7   pp.1941-1944
Publication Date: 2016/07/01
Online ISSN: 1745-1361
DOI: 10.1587/transinf.2016EDL8008
Type of Manuscript: LETTER
Category: Pattern Recognition
Keyword: 
transfer learning,  multiple source transfer,  domain similarity,  manifold assumption,  

Full Text: PDF>>
Buy this Article




Summary: 
Transfer learning extracts useful information from the related source domain and leverages it to promote the target learning. The effectiveness of the transfer was affected by the relationship among domains. In this paper, a novel multi-source transfer learning based on multi-similarity was proposed. The method could increase the chance of finding the sources closely related to the target to reduce the “negative transfer” and also import more knowledge from multiple sources for the target learning. The method explored the relationship between the sources and the target by multi-similarity metric. Then, the knowledge of the sources was transferred to the target based on the smoothness assumption, which enforced that the target classifier shares similar decision values with the relevant source classifiers on the unlabeled target samples. Experimental results demonstrate that the proposed method can more effectively enhance the learning performance.