Owing to labeling training samples are often hard to be collected and subject to a significant amount of human labor or costs, in this paper, a semi-supervised linear subspace learning approach via sparse coding is proposed. Our method can combine labeling samples and unlabeled samples to train a semi-supervised model. First, we train the samples to obtain the structured dictionary and the sparse representation feature; then we group them for more discriminative part (MDP) and less discriminative part (LDP). Finally, we learn the semi-supervised projection matrix P. The experimental results show that the proposed method can improve the classification results on the face image databases extended Yale B and AR datasets.