Link prediction methods based on generalized common neighbor in directed network

Abstract

Link prediction aims to predict missing or future links through currently observed information of network. Existing mainstream methods are mostly applied to undirected network, and some methods designed for directed network ignored the diverse heterogeneous features of common neighbor. For this problem, a generalized common neighbor algorithm was proposed. Firstly, a generalized common neighbor was defined for the directed network. Then the degree of contribution of different structures was measured by the joint edge probability of the directed neighbor isomers, and the existing undirected local similarity index is improved by the new definition, redefining eight kinds of directed similarity indicators based on generalized common neighbor. Experiments on 12 datasets show that proposed method generally improves the performance of existing predictive indicators under two metrics.

Keywords

• directed network
• link prediction
• common neighbor
• heterogeneous fusion