Network embedding has been a popular branch of deep learning, which represents the network information by mapping network nodes to an extended low-dimensional space. According to user co-tag network and social network, we employed network representation learning to extract the representation vector of tagging relationship and social relationship of users respectively, and proposed a novel service recommendation method, using user's representation vector learned to calculate the similar user set and recommend appropriate services to the target users according to the preferences of top-k similar users. To investigate the feasibility of our approach, experiments were carried out on two open data sets, Delicious and Last.FM. The results show that our method outperforms the four benchmarks, with an average improvement of 13% in precision, 18.6% in recall and 13.1% in F-measure. It is also found that when learning user representation, the co-tag relationship between users is as important as the social relationship. Meanwhile, during the process of collaborative recommendation, the number of similar users returned for a target user is suitable in the range from 25 to 30.