Abstract
•MVTL-CP integrates transfer learning and multi-view features into a framework.•The performance of MVTL-CP is theoretically analyzed from two perspectives.•Experimental results confirm the effectiveness of the proposed method.
Transfer learning builds models for the target domain by leveraging the information from another related source domain, in which the distributions of two domains are usually quite distinct. Real-world data are often characterized by multiple representations known as multi-view features. In the multi-view transfer learning field, existing methods aim to address the following two issues. Firstly, due to the distributional difference between the two domains, the classifier trained on the source domain may underperform on the target domain. Moreover, the lack of data from the target domain generally occurs in the training phase. Secondly, how to fully exploit the relations among multiple features is challenging when such multi-view representations emerge in the source and target domains. In this paper, we propose a new coupling loss and self-used privileged information guided multi-view transfer learning method (MVTL-CP). The first issue is addressed by utilizing the weighted labeled data from the source domain to learn a precise classifier for the target domain. Following the consensus and complementarity principles, we tackle the second issue by making the best use of multiple views. Furthermore, we analyze the consistency between views and the generalization capability of MVTL-CP. Comprehensive experiments confirm the effectiveness of our proposed model.