Learning Structural Node Representations on Directed Graphs

Niklas Steenfatt, Giannis Nikolentzos, Michalis Vazirgiannis, Qiang Zhao

Abstract: Many applications require identifying nodes that perform similar functions in a graph. Learning latent representations that capture such structural role information about nodes has recently gained a lot of attention. A state-of-the-art algorithm, struc2vec, generates such representations for the nodes of undirected networks. However, the algorithm is unable to handle directed, weighted networks. In this paper, we present struc2vec++, a generalization of the above algorithm to such types of networks. We evaluate struc2vec++ on real and synthetic networks. We show that taking into account edge directions greatly improves performance. We compare struc2vec++ against a recently proposed algorithm. Although struc2vec++ is in most cases outperformed by the competing algorithm, experiments in a variety of different scenarios demonstrate that it is much more memory efficient and it can better capture structural roles in the presence of noise.