Hierarchical Graph Representation Learning with Structural Attention for Graph Classification

Abstract

Recently, graph neural networks (GNNs) exhibit strong expressive power in modeling graph structured data and have been shown to work effectively for graph classification tasks. However, existing GNN models for predicting graph categories frequently neglect the graph hierarchy information or fail to accurately capture the graph substructure, resulting in significant performance decreases. In this paper, we propose Hierarchical Graph Representation Learning (HGRL), a multi-level framework for capturing hierarchical local and global topological structures to enrich graph representations. In specific, we utilize a structural coarsening module that generates a series of coarsened graphs for an input graph instance, followed by a graph encoder to preserve the local graph structure information. Furthermore, graph convolutional networks are layered to capture high dimensional proximity in graphs, and we incorporate the attention mechanism for entire graph embedding, which enables our framework to focus on critical nodes with significant contribution to the learned low-dimensional graph representations for subsequent graph classification tasks. Experimental results on multiple benchmark datasets demonstrate that the proposed HGRL can substantially improve the classification accuracy and outperform the existing state-of-the-art graph classification approaches.