Large-scale network embedding is to learn a latent representation for each\nnode in an unsupervised manner, which captures inherent properties and\nstructural information of the underlying graph. In this field, many popular\napproaches are influenced by the skip-gram model from natural language\nprocessing. Most of them use a contrastive objective to train an encoder which\nforces the embeddings of similar pairs to be close and embeddings of negative\nsamples to be far. A key of success to such contrastive learnin…