Hyperbolic Attributed Network Embedding with self-adaptive Random Walks

Abstract

Network embedding aims to learn low-dimensional vectors for vertices in complex networks. In real-world systems, nodes in networks are commonly associated with diverse attributes. However, classic approaches generally ignored the implicit relations and hierarchical information introduced by attributes. Motivated by this, we propose a new method named AHANE, short for Adaptive Hyperbolic Attributed Network Embedding, to learn the vertex representations of attributed networks. We perform a biased self-adaptive random walk, generating vertices sequences that can well retain the distribution of vertices in attributed networks. And then propose a novel framework to optimize both the explicit relations (i.e., observed directly connected links between nodes) and implicit relations (i.e., unobserved but transitive links through attributes) by using hyperbolic skip-gram model. We conducted extensive experiments on real datasets related to vertex classification, link prediction and nearest nodes searching. Experimental results on real-world datasets demonstrate the efficiency and effectiveness of AHANE.