Extractive adversarial networks for network embedding

Abstract

Network embedding has attracted more and more researchers recently. Although many algorithms focus on topological information, there exists a disadvantage. Nodes in many real-world network often have their own attributes, which are potentially valuable information, but algorithms focusing on structure ignore these messages, which will decrease the accuracy of the embedding space. To solve this problem, we consider a deep learning model. In this paper, we propose a novel model named Extractive Adversarial Networks for network embedding, which aims to extract the latent space from labels, attributed information and topological structure (e.g. low-order proximity), based on generative adversarial network model in Domain-Adversarial Training of Neural Networks. The goal of EAN is extracting the same features from the topological structure and attributed information according to our extractive network, while distinguishing the extractive features from either the topological structure or attributed information, which becomes an adversarial models. First, we train our extractive network with labels information, which is feeded with attributed information and nodes structure in vector forms. The second step is training the adversarial network to distinguish the embedding representation from either the topological structure or attributed information, where the embedding representation is from the hidden layer in extractive network. We conduct experiments on six real-world datasets, which illustrates that the proposed models outperforms state-of-The-Art embedding algorithms.