Semi-supervised classification via full-graph attention neural networks

Graph neural networks (GNNs) leverage graph convolutions or their approximations to extract features of nodes from graph-structured data. Nevertheless, these methods only combine information from nodes' neighborhoods, without taking into account the impact of other nodes outside neighborhoods. To address the shortcomings, we present full-graph attention neural networks (FGANNs), novel neural network architectures that consider the impact of all nodes when performing self-attention, leveraging masked attention to enable (implicitly) specifying different weights to different nodes in a neighborhood. Under such circumstances, we address several important challenges of spectral-based graph neural net-works simultaneously, and make FGANN readily available to semi-supervised classification problems. Extensive experiments on citation networks offer evidence that the proposed approach outperforms state-of-the-art methods. (c) 2021 Elsevier B.V. All rights reserved.