Research on a link-prediction method based on a hierarchical hybrid-feature graph

Abstract

Entities in the real world are often interconnected, forming heterogeneous information networks. Link-prediction is a necessary technique for predicting the existence of unobserved or future links in heterogeneous information networks. It is useful to make users better understand the generation and evolution of networks. However, current techniques lack the effective fusion of multiple features, often leading to nonsensical results. Also, it is difficult for them to adapt to the heterogeneity and dynamics of heterogeneous-information networks. In this paper, we present a hierarchical-hybrid-feature-graph (HHFG) model by fully considering structural, semantic, and time features. Also, an HHFG based link-prediction algorithm is proposed to effectively guarantee accuracy. On one hand, it performs a random walk on HHFG based upon hybrid features. On the other hand, parameters such as feature weights and transition coefficients are learned by the gradient-descent method. The experiments demonstrate the feasibility and effectiveness of our key techniques.