In action recognition, in order to obtain high performance classifiers, it is necessary to feed the training algorithms enough labeled data. Since labeling is a very expensive task, it is important to develop approaches which can efficiently reuse existing labeled data. In this work, we consider the task of utilizing labeled data from one dataset (source dataset) to train action classifiers for data from another completely unlabeled dataset (target dataset). We propose a novel approach for such a task by extending the well-known self-training algorithm to including data selection or feature selection. The superior of our approach has been confirmed by benchmark datasets.