Recently researchers have proposed activity recognition methods based on adapting activity knowledge obtained in previous spaces to a new space. Adapting activity knowledge allows us to quickly recognize activities in a new space using only small amounts of unlabeled data. However, adapting from dissimilar spaces not only does not help the recognition task, but might also lead to degraded recognition accuracy. We propose a method for automatically selecting the most promising source spaces among a number of available spaces. Our approach leads to a scalable and quick solution in real world, while minimizing the negative effects of adapting from dissimilar sources. To evaluate our algorithms, we tested our algorithms on eight real smart home datasets.