<i>struc2vec</i>

Abstract

Structural identity is a concept of symmetry in which network nodes are\nidentified according to the network structure and their relationship to other\nnodes. Structural identity has been studied in theory and practice over the\npast decades, but only recently has it been addressed with representational\nlearning techniques. This work presents struc2vec, a novel and flexible\nframework for learning latent representations for the structural identity of\nnodes. struc2vec uses a hierarchy to measure node similarity at different\nscales, and constructs a multilayer graph to encode structural similarities and\ngenerate structural context for nodes. Numerical experiments indicate that\nstate-of-the-art techniques for learning node representations fail in capturing\nstronger notions of structural identity, while struc2vec exhibits much superior\nperformance in this task, as it overcomes limitations of prior approaches. As a\nconsequence, numerical experiments indicate that struc2vec improves performance\non classification tasks that depend more on structural identity.\n