Link prediction in dynamic networks aims to predict edges according to historical linkage status. It is inherently difficult because of the linear/non-linear transformation of underlying structures. The problem of efficiently performing dynamic link inference is extremely challenging due to the scale of networks and different evolving patterns. Most previous approaches for link prediction are based on members' similarity and supervised learning methods. However, research work on investigating hidden patterns of dynamic social networks is rarely conducted. In this paper, we propose a novel framework that incorporates a deep learning method, i.e., temporal restricted Boltzmann machine, and a machine learning approach, i.e., gradient boosting decision tree. The proposed model is capable of modeling each link's evolving patterns. We also propose a novel transformation for input matrix, which significantly reduces the computational complexity and makes our algorithm scalable to large networks. Extensive experiments demonstrate that the proposed method outperforms the existing state-of-the-art algorithms on real-world dynamic networks.