DynGAN: Generative Adversarial Networks for Dynamic Network Embedding

Embedding large graphs in a low-dimensional space has proven useful in various applications. However, there is a limited focus on real-world networks that are dynamic in nature and continuously evolving with time. In this paper, we propose a novel adversarial algorithm to learn representation of dynamic networks. We leverage generative adversarial networks and recurrent networks to capture temporal and structural information. We conduct extensive experiments on the task of graph reconstruction, link prediction and graph prediction. Experimental results demonstrate consistent, stable, and better results against state-of-the-art methods in many cases.