Cross-regression for multi-view feature extraction

Abstract

The traditional multi-view feature extraction (MvFE) method usually seeks a latent common subspace where the samples from different views are maximally correlated. Recently, the regression-based method has become one of the most effective feature extraction methods. However, the existing regression-based methods are only suitable for single-view cases. In this paper, we firstly propose a new MvFE method named as cross-regression for MvFE (CRMvFE). CRMvFE designs a novel cross-regression regularization term to discover the relationship between multiple views in the original space, and simultaneously obtains the low-dimensional projection matrix for each view. Furthermore, inspired by the robustness of L2,1-norm, we also propose a robust CRMvFE (RCRMvFE) and an iterative algorithm to find the optimal solution. Theoretical analysis of the convergence and the relationship with CRMvFE demonstrate the effectiveness of the proposed RCRMvFE. Experiments on datasets show that the proposed CRMvFE and RCRMvFE have better performance than other related methods.