Geometric-aware Graph Attention Networks

Graph Neural Networks (GNNs) have recently drawn wide public attention due to the tremendous expressive power for graph analytics. Most existing GNN models convey messages among nodes solely based on the shallow topology information (e.g., connections), while deeper graph structural signals are rarely explored. Our motivation lies in capturing the geometric consistency in terms
of edges and angles to provide extra geometrical learning guidance. In this paper, we propose a novel Geometric-aware Graph Attention Networks (GeomGAT) model for preserving the geometric characteristics from the input graph under two elaborate topological constraints. We introduce a multi-view attention mechanism to encode the geometric information adequately. In order to alleviate the challenge of scarce topological information, we further propose a novel contrastive learning-based framework with geometric augmentation. Comprehensive experiments conducted on multiple publicly available datasets demonstrate the superiority of our proposal.