There exist several problems in existing graph-based semi-supervised learning (GSSL) methods such as model parameters sensitiveness and insufficient discriminative information in data space, etc. To address those issues, this paper proposes a sparse approximated nearest feature space embedding label propagation (SANFSP) algorithm, which is inspired by both ideas of nearest feature space embedding and that of sparse representation. SANFSP first sparsely reconstructs data from original space using its feature space embedding projection images, and then measures the similarity between original data and its sparse approximated nearest feature space embedding projection points, thus proposing a sparse approximated nearest feature space embedding regularizer. At last, SANFSP complets label propagation procedure by using classical label propagation algorithm. The study also derives an easy way to extend SANFSP to out-of-sample data. Promising experimental results are obtained on several toy and real-world classification tasks such as face recognition, visual object recognition and digit classification.