In order to make full use of user’s tagging and social relationship information and improve the accuracy of recommendation results in service recommendation, we propose a new recommendation method based on heterogeneous user network embedding. It maps a user node to a lowdimensional vector, and then utilizes the obtained user vector to carry out collaborative recommendation. We verify the method on Delicious, a public dataset. Experimental results show that our method outperforms the two existing methods, and the recommended accuracy is increased by 18.1% and 16.6% respectively. Furthermore, the results also suggest that the direct relationship between nodes is as important as the "friends of friends" relationship in representing user node structural information when learning the user representation vector. At the same time, it is most suitable to return top 25 similar users for the target user in the recommendation process.