In this paper, we introduce and study a novel unsupervised domain adaptation (DA) algorithm, called latent subspace sparse representation based domain adaptation, based on the fact that source and target data that lie in different but related low-dimension subspaces. The key idea is that each point in a union of subspaces can be constructed by a combination of other points in the dataset. In this method, we propose to project the source and target data onto a common latent generalized subspace which is a union of subspaces of source and target domains and learn the sparse representation in the latent generalized subspace. By employing the minimum reconstruction error and maximum mean discrepancy (MMD) constraints, the structure of source and target domain are preserved and the discrepancy is reduced between the source and target domains and thus reflected in the sparse representation. We then utilize the sparse representation to build a weighted graph which reflect the relationship of points from the different domains (source-source, source- target, and target-target) to predict the labels of the target domain. We also proposed an efficient optimization method for the algorithm. Our method does not need to combine with any classifiers and therefore does not need train the test procedures. Various experiments show that the proposed method perform better than the competitive state of art subspace-based domain adaptation.