Attention aware edge-node exchange graph neural network

Abstract

An attention aware edge-node exchange graph neural network (AENN) model was proposed, which used edge-node switched convolutional graph neural network method for graph encoding in a graph structured data representation framework for semi supervised classification and regression analysis.AENN is an universal graph encoding framework for embedding graph nodes and edges into a unified latent feature space.Specifically, based on the original undirected graph, the convolution between edges and nodes was continuously switched, and during the convolution process, attention mechanisms were used to assign weights to different neighbors, thereby achieving feature propagation.Experimental studies on three datasets show that the proposed method has significant performance improvements in semi-supervised classification and regression analysis compared to existing methods.

Keywords

• graph neural network;message passing;attention mechanism;hypergraph;line graph