A Graph Embedding Method Based on Opinion Dynamics

Abstract

∗Graph (a.k.a. network), as an important form of data representation, exists widely in real scenes. Effective graph analysis helps users to understand the information hidden behind the data, which will benefit the following tasks (such as node classification, link prediction). Graph embedding is a very effective method to solve the problem of graph analysis. It maps the graph data into the low-dimensional space, and retains the structure and attribute information of the graph to the maximum extent. Graph embedding has become a hot topic in recent years.We propose a neural network graph embeddingmethod based onOpinionDynamics, called Graph Opinion Dynamics Networks (GODNs), which is a neural network architecture that adopts a new information aggregation strategy to process structured data, leveraging attentional mechanism and trusted neighours convergence to address the shortcomings of prior methods based on graph convolutions. Node embeddings are updated through multiple rounds of aggregation of trusted neighbors with different weights, and the rules are set to achieve convergence state. For transductive semi-supervised learning problems, we divide the datasets into sparse networks and dense networks. Our GODNs models have matched the results across the sparse networks (the Cora, Citeseer and Pubmed citation network datasets) and achieved a great improvement over dense networks (amazonphoto, amazon-computer and coauthor-cs electronic commerce datasets).