Dual-view hypergraph neural networks for attributed graph learning

Abstract

Graph embedding analyzes network data by learning the vector representation of each vertex in the network, and has attracted widespread attention in recent years. In many real-world networks, considering the topology and attributes of nodes comprehensively has potential value in achieving more effective graph embedding. However, the existing methods do not fully utilize and integrate these two kinds of information. The main challenges include three aspects: sparseness and noise in structure information; insufficient modeling of non-linear relationship between nodes in attribute semantic space; heterogeneity of structure and attribute information. To address these issues, based on graph neural network and hypergraph, we propose a Dual-view HyperGraph Neural Network (DHGNN) model for attributed graph learning. First, we unify the expression form of different information sources of nodes by hypergraph, and construct dual hypergraphs according to topology and attributes of nodes. Secondly, we propose a dual-view hypergraph neural network for graph embedding. The central idea is that we model and integrate different information sources by shared and specific hypergraph convolutional layer, and use the attention mechanism to adequately combine dual node embeddings. Finally, we train the model through semi-supervised node classification task. Extensive experiments have been carried out on four real world public datasets, demonstrating the performance of the proposed model DHGNN has always been superior to that of state-of-the-art graph embedding methods.