Role-Oriented Graph Auto-encoder Guided by Structural Information

Roles in a complex network usually represent the local connectivity patterns of nodes, which reflect the functions or behaviors of corresponding entities. Role discovery has great meaning for understanding the formation and evolution of networks. While the importance of role discovery in networks has been realized gradually, a variety of approaches of role-oriented network representation learning are proposed. Almost all the existing approaches are dependent on manual high-order structural properties which are always fragmentary. They suffer from unstable performances and poor generalization ability, because their hand-craft structural features sometimes miss the characteristics of different networks. In addition, graph neural networks (GNNs) have great potential to automatically capture structural properties, but it is hard to be given the rein to for the difficulty of designing role-oriented unsupervised loss. To overcome these challenges, we provide an idea that leverage low-dimensional extracted structural features as guidance to train graph neural networks. Based on the idea, we proposed GAS, a novel graph auto-encoder guided by structural information, to learn role-oriented representations for nodes. Results of extensive experiments show that GAS has better performance than other state-of-the-art approaches.