Hierarchical Message-Passing Graph Neural Networks

09/08/2020
by   Zhiqiang Zhong, et al.
30

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.

READ FULL TEXT

Please sign up or login with your details

Forgot password? Click here to reset