A robust multi-source transfer classification method based on belief functions for cross-domain pattern recognition

Abstract

In pattern recognition with few or no labeled data, domain adaptation techniques are often used to transfer knowledge in the source domain to help build classification models in the target domain. The effective combination of complementary information in multiple source domains usually can further improve the classification accuracy. To this end, we present a Robust Multi-Source Transfer (RMST) classification method consisting of two-step fusion of classifiers to extract the useful information in each source domain as much as possible, and to effectively combine the complementary information using belief functions in different source domains. The first step is to accurately predict the pseudo labels when computing conditional distribution, in order to learn a robust new feature representation, which is used to fuse diverse classifiers learnt by patterns for reliable soft classification in the second step of our algorithm. Furthermore, the soft classification results yielded by the assistance of different source domains are combined by belief functions with the new weighting factors taking into account the distribution discrepancy and classifier’s performance. The effectiveness of RMST was evaluated with respect to a variety of advanced methods, and the experimental results show that RMST can significantly improve the classification accuracy.