Recommender systems always offer the most suitable goods for users by adopting collaborative filtering or content-based technology based on historical ratings in one domain. However, sometimes we may suffer from item cold start problem in a new domain, especially in Cyber-Physical Systems (CPS). To alleviate this problem, many recommendation models have been proposed to transfer one domain's knowledge to another domain by using transfer learning algorithm. In this paper, we propose a new cross-domain recommendation algorithm which is divided into two stages. In the first stage, we apply the TrAdaBoost algorithm to select some items which are worthy of being recommended to users in the target domain. Then in the second stage, we adopt the nonparametric pairwise clustering algorithm to make a decision whether to recommend an item to a group of users or not. We not only make a classification for the target domain items but also find the recommended or not recommended customer groups for one item through the two stages. Experiments on real world data sets demonstrate that our proposed method performs better than other algorithms for the cross-domain recommendation task.