Dynamic information network is a new challenging problem in the field of current complex networks. The evolution of dynamic networks is temporal, complex and changeable. Structure is the basic characteristics of the network, and is also the basis of network modeling and analysis. The study of the network structure evolution is of great importance in getting a comprehensive understanding of the behavior trend of complex systems. This paper introduces "role" to quantify the structure of dynamic network and proposes a role-based model. To predict the role distributions of dynamic network nodes in future time, the presented framework views role prediction as a multi-target regression problem, extracts properties from historical snapshot sub-network, and predicts the future role distributions of dynamic network nodes. The paper then proposes a multi-target regression based role prediction (MTR-RP) method for dynamic network. This method not only overcomes the drawback of the existing methods which operate on transfer matrix while ignoring the time factor, but also takes into account of possible dependencies between multiple forecast targets. Experiments results show that MTR-RP has better and more stable prediction capability compared with the existing methods.