LAUSR

Multi-view feature learning aims at improving the performances of learning tasks, by fusing various kinds of features (views), such as heterogeneous features and/or homogeneous features. Current leading multi-view feature learning approaches usually learn features in each view separately while not uncovering shared information from multiple views. In this paper, we propose a multi-view feature learning framework, which can simultaneously learn separate subspace for each view and shared subspace for all the views, respectively; specifically, the separate subspace for each view can preserve the particular information within this view, meanwhile, the shared subspace can capture feature correlation among multiple views. Both the particularity and communality are essential for classification. Furthermore, we relax the labels of training samples within the concatenated subspaces, thus resulting in the retargeted least square regression (LSR) classifier. The transformation matrices tailored for each subspace within the corresponding view and the label relaxed LSR classifier are jointly learned in a unified framework, based on an efficient alternative optimization manner. Extensive experiments on four benchmark data sets well demonstrate the superiority of the proposed method, which has led to better performances than compared counterpart methods.