Hierarchical Message-Passing Graph Neural Networks

Graph Neural Networks (GNNs) have become a promising approach to machine learning with graphs. Since existing GNN models are based on flat message-passing mechanisms, two limitations need to be tackled. One is costly in encoding global information on the graph topology. The other is failing to model meso- and macro-level semantics hidden in the graph, such as the knowledge of institutes and research areas in an academic collaboration network. To deal with these two issues, we propose a novel Hierarchical Message-Passing Graph Neural Networks framework. The main idea is to generate a hierarchical structure that re-organises all nodes in a graph into multi-level clusters, along with intra- and inter-level edge connections. The derived hierarchy not only creates shortcuts connecting far-away nodes so that global information can be efficiently accessed via message passing but also incorporates meso- and macro-level semantics into the learning of node embedding. We present the first model to implement this hierarchical message-passing mechanism, termed Hierarchical Community-aware Graph Neural Network (HC-GNN), based on hierarchical communities detected from the graph. Experiments conducted on eight datasets under transductive, inductive, and few-shot settings exhibit that HC-GNN can outperform state-of-the-art GNN models in network analysis tasks, including node classification, link prediction, and community detection.