Multi-View Graph Regularized Discriminant Analysis

In computer vision and multimedia search, it is common to use multiple descriptions from different views to represents an object. As one of the popular multi-view learning methods, Multi-view Linear Discriminant Analysis (MLDA) was proposed to maximize discrimination in each view and correlation between two views simultaneously through combining CCA and LDA. However, MLDA dues not utilize the local discriminant information between classes. In this paper, we propose a novel supervised learning method, Multi-view Graph Regularized Discriminant Analysis (MGRDA), to offset the shortage of MLDA. MGRDA constructs three adjacency graphs to model the global discriminant structure, local discriminant structure and intrinsic structure in each view, respectively, in order to enhance the inter-class discrimination. Then these three graph regularizers arc integrated with Discriminant-CCA so as to preserve both the correlation and class structures along each pair of views. Experimental results on benchmark databases verify the effectiveness of the proposed method, indicating the benefits from mining the information of inter-class local structures in multi-view data.