Collaborative weighted multi-view feature extraction

Abstract Most of the current multi-view feature extraction methods mainly consider the consistency and complementary information between multi-view samples, therefore have some drawbacks. They ignore the manifold structure of the single-view itself, and also ignore the differences among the similarities between any two views when the number of views is greater than two, because of assigning the same weight to them. In this paper, we propose a novel multi-view feature extraction method termed as collaborative weighted multi-view feature extraction or CWMvFE. Here the local collaborative representative (LCR) method is utilized to preserve the local correlation in between-view and within-view respectively. Furthermore, it realizes that less similar view pairs should share more consistency and complementary information, where Jensen Shannon divergence is used to reflect the similarity between different view pairs. Therefore, the proposed CWMvFE not only preserves the local correlation in multi-view, including local correlation in both between-view and within-view, but also explores the differences in similarities between different view pairs. Experiments on four image datasets demonstrate that CWMvFE has better performance than other related methods.