Cross-Database Micro-Expression Recognition (CDMER) aims to develop the Micro-Expression Recognition (MER) methods that satisfy different conditions (equipment, subjects, and scenes) in practical application, i.e., the MER method of strong domain adaption ability. CDMER faces two obstacles: 1) the severe feature distribution gap between the training and test databases and 2) the feature representation bottleneck for micro-expression (ME) such local and subtle facial expressions. To solve these obstacles, this paper proposes a novel Transfer Group Sparse Regression method, namely TGSR, which seeks and selects those salient facial regions to 1) promote a more precise measurement of the difference between source and target databases by the operation in the feature level to alleviate their difference better, and to 2) improve the extracted hand-crafted feature to be more effective and explicable for better MER. We use two public micro-expression databases, i.e., CASME II and SMIC, to evaluate the proposed TGSR. Experimental results show that TGSR learns the discriminative feature and outperforms most state-of-the-art subspace-learning-based domain adaption methods for CDMER.