Learning Network Representation Via Ego-Network-Level Relationship

Network representation, aiming to map each node of a network into a low-dimensional space, is a fundamental problem in the network analysis. Most existing works focus on the self-level or pairwise-level relationship among nodes to capture network structure. However, it is too simple to characterize the complex dependencies in the network. In this paper, we introduce the theory of the ego network and present an ego-network-level relationship. Then a deep recurrent auto-encoder model is proposed to preserve the complex dependencies in each ego network. In addition, we present two strategies to solve the sparsity problem. Finally, we conduct extensive experiments on three real datasets. The experimental results demonstrate that the proposed model can well preserve network structure and learn a good network representation.