Multi-view laplacian least squares for human emotion recognition

Abstract

Human emotion recognition is an emerging and important area in the field of human–computer interaction and artificial intelligence, which has been more and more related with multi-view learning methods. Subspace learning is an important direction of multi-view learning. However, most existing subspace learning methods could not make full use of both category discriminant information and local neighborhood information. As a typical subspace learning method, partial least squares (PLS) performs better and more robustly than many other subspace learning methods, because PLS is optimized with iteration method. However, PLS suffers from linear relationship assumption and two-view limitation. In this paper, a new nonlinear multi-view laplacian least squares (MvLLS) is proposed. MvLLS constructs a global laplacian weighted graph (GLWP) to introduce category discriminant information as well as protects the local neighborhood information. Optimized with iteration method, MvLLS is a multi-view extension of PLS. The proposed method has great extendibility and robustness. To meet the requirements of large-scale applications, weighted local preserving embedding (WLPE) is proposed as the out-of-sample extension of MvLLS, basing on the idea of maintaining the manifold structures of original space. Finally, the proposed method is verified on three multi-view emotion recognition tasks, the experiment results validate the effectiveness and robustness of MvLLS.