Multi-view subspace learning via bidirectional sparsity

Abstract

With the improvement of multi-view data collection technology, multi-view learning has become a hot research area. How to deal with diverse and complex data is one of the challenging problems in multi-view learning. However, it is hard for traditional multi-view subspace learning methods to find an effective subspace dimension and deal with outliers simultaneously. In this paper, we propose a novel method, named as Multi-view Subspace Learning via Bidirectional Sparsity(SLBS), which is effective to overcome the above difficulties and learn a better representation. Specifically, we divide the shared subspace into two parts. One is a row sparse matrix to do a secondary extraction of features and the other is a column sparse matrix to reduce the influence of outliers. The proposed model is a non-convex problem which is difficult to be solved. To address this problem, we develop an efficient algorithm and analyze its convergence and computational complexity. Finally, compared with other multi-view subspace learning methods, the extensive experimental results on real-world datasets present the effectiveness of our SLBS.