In the existing study of user roles, many scholars have defined the number and characteristics of roles，which have achieved good results in a particular dataset. But there are two problems: 1) the generality is poor,i.e., it must be reanalyzed if the dataset has been replaced; 2) in the real world, user's behavior and relationships are complicate and user roles are varied． So it's very difficult to describe and identify them with the artificial definition. So, this article proposed a user role found algorithm based on the tensor decomposition model. This algorithm can not only set the number of roles automatically, but also reflect the behavior characteristics of the role in specified period of time. Furthermore, this article extended user role to community roles and raised a community evolution analysis method based on the distance of community roles and the node overlapping. The experiment results indicate that the behavior characteristics of the identified roles are consistent with the fact, and the community evolution analysis method proposed has better effect than comparison algorithms.