Multi-granularity Network Representation Learning Based on Game Theory

Hang Shu, Qun Liu, Shuyin Xia

Abstract: The study of embedding large information networks into low-dimensional vector spaces has been going on for several years. However, current feature learning methods are not sufficient to simultaneously capture the structural identity and homophily of nodes in the network. We propose Multi-granularity Network Representation Learning (MGNRL), a novel and flexible framework to learn the latent representations of nodes. Firstly, nodes are divided into different granular layers according to their importance by MGNRL, and the neighborhood structures of nodes are considered to encode their similarities. Then a random walk procedure is designed to explore nodes on different granularity layers, and the public goods game theory is introduced to adjust the probability of random walk continuously. MGNRL overcomes the limitations of existing methods by simultaneously capturing the structural identity and homophily of network nodes. Experiments show that MGNRL performs well in many tasks such as visualization, node classification and link prediction.