Dynamic Network Embedding with Random Points on the Hypersphere

Dynamic network embedding has recently attracted significant attention due to the challenges of network evolvement. Although previous network representation methods have proposed some static graph embedding algorithms but ignore rapid node adding situation in dynamic network. In this paper, we study the problem of learning new-added node embeddings in dynamic network. We propose a new methodology to compute node embeddings with the help of the evolvement of network structure. Firstly, we initialize the original node embeddings with static methods. Then we generate a delta vector on the hypersphere to obtain the new-added node embeddings. Additionally, our method boost the progress of network embedding by using the disappearing node at the stage of negative sampling. Finally, we evaluate extensive experiments on dynamic network data sets. Compared with previous static methods, the experimental results show that our algorithm is able to achieve better time cost and obtain excellent results in various machine learning tasks including link prediction and node classification task.