I. Introduction
Graph structures, e.g., citation networks and social networks, are ubiquitous and fast-growing in the real world. Network representation learning (NRL) can map the semantic similarity of graph vertices into a low-dimensional vector space where similar vertices are assigned to the nearby areas [1]. The learned representations are useful for the subsequent applications, such as vertex classification [2], link prediction [3], and data visualization [4]. As demonstrated in the above applications, more discriminative representations of vertices would benefit for the better performance of the downstream tasks. Thus, the key to the success of the downstream applications is learning discriminative representations of vertices.