Semi-supervised classification of joint graph convolution and hypergraph convolution

In recent years, hypergraphs have high-level modeling capabilities in real-world scenarios, and hypergraph deep learning methods have been used for semi-supervised classification tasks on hypergraph data. However, there are some shortcomings existing hypergraph convolution neural networks. The current hypergraph neural networks introduce noise in the process of multi-layer convolution node neighborhood expansion, which make it difficult to extract discriminative features. There is a relatively high model complexity in the traditional multi-channel convolution process. To solve these problems, the joint hypergraph convolution and graph convolution neural network was proposed. First, hypergraph convolution was used to extract high-order correlation information on the original hypergraph data. Second, the hypergraph was constructed into a graph, and the obtained high-order node features were combined with the graph structure. Finally, graph convolutions were exploited to aggregate local information within the first-order correlation neighborhood of the nodes. The results of node classification experiments on three citation networks show that the proposed algorithm can achieve higher classification accuracy than existing algorithms, and the parameter amount and training time are about half of those of traditional multi-channel hypergraph neural networks.