This paper focuses on the research of cross-database micro-expression recognition, in which the training and test micro-expression samples belong to different microexpression databases. Mismatched feature distributions between the training and testing micro-expression feature degrade the performance of most well-performing micro-expression methods. To deal with cross-database micro-expression recognition, we propose a novel domain adaption method called Transfer Group Sparse Regression (TGSR). TGSR learns a sparse regression matrix for selecting salient facial local regions and the corresponding relationship of the training set and test set. We evaluate our TGSR model in CASME II and SMIC databases. Experimental results show that the proposed TGSR achieves satisfactory performance and outperforms most state-of-the-art subspace learning-based domain adaption methods.