Inductive Structural Role Embedding on Large-scale Graphs

Zheyu Zhang, Yuanhao Huang, Jie Liu

Abstract: Graph embedding methods have been proposed to extract structural identities in a graph, but most of the existing structural role embedding methods are transductive and the embedding cannot be generalized to unseen nodes. Here we introduce InSuRE, an inductive method to embed nodes' structural roles. Instead of leveraging a diffusion process on the entire graph, we characterize a local diffusion kernel with two learnable parameters, the local neighborhood radius and corresponding diffusion scale. With the two parameters, the embedding of unseen nodes can be efficiently generated based on their neighborhood topology. InSuRE is computationally efficient, provides discriminative structural features to improve GNN's expressive power, and outperforms baseline methods in empirical experiments.