A new multi-source Transfer Learning method based on Two-stage Weighted Fusion

Transfer Learning (TL) as an effective tool to solve the classification problem with few or no labeled training data usually confronts with the multiple source domains issue. Fusion of complementary knowledge in different source domains is expected to improve the classification accuracy in the target domain. For this purpose, we present a new Multi-source transfer learning method via Two-stage Weighted Fusion (MTWF). In MTWF, a credal classification rule is first developed to preserve the local imprecise information as much as possible because it could be difficult to commit some patterns to one precise class when using TL techniques. Then, multiple credal classification results characterized by mass function values produced by patterns in each source domain under different feature spaces are locally integrated to extract useful information in each source domain. The locally integrated credal classification results in different source domains are regarded as multiple sources of evidence for global combination using belief functions to make the class decision, and they are first discounted via the weights estimated by domain-consistency. Whereas, there exist some conflicts among these discounted credal classification results, so we proposed to further discount them via Belief Jensen-Shannon (BJS) divergence for the small conflicts. MTWF was compared with a variety of advanced methods, and the experiment results demonstrate that MTWF can significantly improve the classification accuracy. (c) 2022 Elsevier B.V. All rights reserved.