Graph Autoencoder Combined with Attribute Information in Graph

Abstract

Network embedding is an important method to learn the low-dimensional representation of nodes. It can be used to reduce the dimension of the network for downstream tasks. Many existing embeddings learn the representation only based on the reconstruction of the topological structure, yet nodes with attributes can provide important information in many real networks. Thus, learning attributed graph embedding should not only be based on the reconstruction of topological structure but also node features. In this paper, we propose a reasonable and interpretable graph autoencoder based on structure and attributes. This model can also be extended in both supervised and unsupervised tasks. Experiments including link prediction, clustering, and graph visualization have verified the effectiveness of our proposed GAE model.