Abstract
In this paper, we present DONE, a novel framework for learning networks representations, which fundamentally supports numerous network analytic tasks such as node classification, clustering, and visualization. Most existing network embedding methods are unable to efficiently scale for large networks and usually suffer from performance issues. In addition, these methods cannot efficiently leverage the vertex labels that are usually very scarce in real-world networks. Our framework provides a powerful way to generate representations for network vertices without relying on user-defined heuristics for manual feature extraction. We present a deep autoencoder model to generate low-dimensional feature representations by learning network reconstruction and semi-supervised classification tasks. We propose a novel greedy algorithm based on vertex domination and centrality concepts to simplify networks while preserving network topology and community structure, namely GreedyNet. In addition, we propose a novel sampling approach to estimate the labels of unlabeled vertices and adopt it to learn superior embedding. Using network dominating vertices, our approach enhances the generalization and scalability of network embedding through simplifying the underlying network. To the best of our knowledge, we are the first to propose the idea of using network dominating vertices to enhance network embedding. The experimental results show that our method outperforms the state-of-the-art methods.