Linear Regression Classification (LRC) based face recognition achieves high accuracy while being highly efficient. As with most other linear-subspace-based methods, the faces of a subject are assumed to reside on a linear manifold; however, where occlusion or disturbances are involved, this assumption may be inaccurate. In this paper, a manifold-learning procedure is used to expand on conventional LRC by excluding faces not fitting the original assumption (of linearity), thereby localizing the manifold subspace, increasing the accuracy over conventional LRC and reducing the number of faces for which the regression must be performed. The algorithm is evaluated using two standard databases and shown to outperform the conventional LRC.