Multi-view multi-label learning with double orders manifold preserving

In multi-view multi-label learning, each instance has multiple heterogeneous views and is marked with a collection of non-exclusive discrete labels. This type of data is usually subject to dimensional catastrophe. Previous multi-view multi-label works look for a low-dimensional shared subspace to tackle this problem. However, these methods ignore the global structural information of the original feature space during dimension reduction. In this paper, we propose Multi-view Multi-label learning with Double Orders Manifold Preserving (MMDOM). MMDOM utilizes manifold preserving constraint to guide the formation of low-dimensional shared subspace. To obtain exact manifold preserving, the first-order and the second-order similarity matrices are both introduced to explore the local and global structural information of the original feature space. Experiments on various benchmark datasets demonstrate the superior effectiveness of MMDOM against state-of-the-art methods.