Multi-view Discriminative Manifold Embedding for Pattern Classification

While many dimensionality reduction algorithms have been proposed in recent years, most of them are designed for single view data and cannot cope with multi-view data directly. Dimensionality reduction algorithms in recent ten years, both in theory and application have great breakthrough. In the face of dozens, hundreds or even thousands of dimension by dimension reduction to the data from high dimensional space to a low dimensional space and extract the essential characteristics of low dimensional data. In many real-world pattern applications such as face recognition, multiple feature descriptors can provide complementary information in characterizing image from different viewpoints. Motivated by this concern, we propose a new multi-view discriminative manifold embedding (MDME) method for classification by making use of intra-class geometry and inter-class marginal information as well as complementary information of multiple feature representations. Experimental results on face recognition demonstrate the effectiveness of the proposed algorithm.