<p>Multi-view subspace learning is an important paradigm to reduce gaps among multiple heterogeneous features. Existing multi-view subspace learning methods mostly focus on shared features from all views via consistency, and so they do not specifically explore any view-specific information but suppress it. A few studies recently exploited both shared and view-specific features, but they are restricted to very limited and specific learning tasks such as clustering or to homogeneous features. In this paper, we propose a new multi-view subspace learning framework to model shared and view-specific features through data reconstruction perspective. In the new framework, not only each view can be reconstructed from both shared and its view-specific features, but also the projections that produce the shared and view-specific features are mutually orthogonal. Two models are instantiated from the proposed framework with customized regularization for unsupervised and supervised learning. A novel optimization method is proposed to solve resulting challenging optimization problems with theoretically guaranteed convergence. Extensive experiments on nine multi-view data sets demonstrate that our proposed models produce superior learning results to existing methods.</p>