Generative Adversarial Graph Representation Learning in Hyperbolic Space

Abstract

Representation learning can provide a compact representation of features. There are a large number of representation learning methods which have been successfully applied to feature learning of graph structured data. However, most of the existing graph representation learning methods does not consider the latent hierarchical structure of the data. They mainly focus on high-dimensional Euclidean space for learning. Recent studies have shown that graph structured data are suitable for being embedded in hyperbolic space and that hyperbolic space can be naturally equipped to model hierarchical structures where they outperform Euclidean embeddings. Therefore, we comprehensively consider the hierarchical structure characteristics of graphs, and learn the vector representation of nodes in hyperbolic space by using the principle of generative adversarial learning. By using two models to simultaneously capture hierarchy and similarity and let them compete with each other, the performance of learning is alternately boosted. In this paper, node classification, link prediction and visualization are performed on multiple public datasets. The results show that this method performs well in multiple tasks.