With the development of online media, in the online virtual world, the relationship between people has formed a kind of social network, and these relationships will change over time, called dynamic social network. In order to deal the network as a graph and consider its time characteristics, we set the time point as a fixed interval and divide a series of snapshots into a dynamic social network to observe the connection trend of nodes and predict which nodes will appear or disappear in the graph. This process is called "Dynamic Network Link Prediction" (DNLP), which means that we can infer a specific group of people who will establish a relationship with the target user based on their historical behavior. We propose a new model based on Generative Adversarial Network (GAN), called Soc-GAN, combined with Long Short-Term Memory (LSTM), to do the link prediction of dynamic social networks and deal with the long-term prediction tasks for capture the relevance of vectors between the sequences and classify, distinguish whether the generated prediction snapshot is similar to real data. At the same time, it also has a more robust ability to predict the links that are going to appear or disappear in the next network graphs.