In this paper, we propose a patch based semi-supervised linear regression (PSLR) approach to address single sample per person (SSPP) problem in face recognition, which takes full use of the unlabeled probe samples to learn facial variation information. We partition each face image into several overlapped patches, where each patch corresponds to a mapping matrix of regression model. Then, mapping matrix can be further adjusted to describe facial variation information by mapping unlabeled patches to the equidistant point [1, 1, …1]T that has no identification information. The solution of mapping matrices of each path are incorporated into a global objective function, which utilizes l2,1-norm minimization to reduce the influence of noise and improve discriminability of mapping matrices. Finally, the classification results of all patches are aggregated by voting. Experimental results on three public face databases not only demonstrate the effectiveness of our approach but also show the robustness to expression, illumination, occlusion and time variation.