HireGC: Hierarchical inductive network representation learning via graph coarsening

Shu Zhao, Ci Xu, Ziwei Du, Yan-ping Zhang, Zhen Duan, Jie Chen

Abstract: Inductive network representation learning utilizes node feature information to efficiently generate embeddings for unseen data, achieving significant success in learning vector representations for new nodes within a network. However, most existing methods focus primarily on modeling the information from neighboring nodes of new nodes, limiting the model’s ability to capture the broader network context. In this paper, we propose a Hierarchical inductive network representation learning model based on Graph Coarsening (HireGC), which effectively captures richer hierarchical network information and employs a multi-level decomposition approach for structured network analysis. Specifically, HireGC uses an iterative coarsening algorithm to generate network structural information across multiple layers, then merges this multilayer network information through node-level mapping and group-level inheritance mechanisms to produce high-quality vector representations for new nodes. Unlike existing methods that only consider neighboring nodes, our approach captures the multi level structural information of the entire network, enabling the learning of more effective inductive embeddings from a broader perspective. Extensive experiments on six benchmark graph datasets demonstrate the superiority of HireGC over other state-of-the-art methods, highlighting the crucial role of hierarchical network information in inductive network representation learning.