The invention provides a link prediction model building method. The link prediction model comprises a time sequence restricted Boltzmann machine model and a gradient enhanced decision tree model. Themethod comprises steps: a large amount of network data is grasped from Internet or other multimedia, the network data are preprocessed, the network data are divided to historical data and available data, the network data are inputted to the time sequence restricted Boltzmann machine model, and model parameters are trained; and network topology characteristics of a network data node pair are extracted, a characteristic set is formed and inputted to the gradient enhanced decision tree model, and model parameters are trained. The link prediction model comprises the time sequence restricted Boltzmann machine model and the gradient enhanced decision tree model which are well trained. The link prediction model is built based on the method. The invention also provides a link prediction method, which can predict all links in the next state of the network.