Multi-head Attention Induced Dynamic Hypergraph Convolutional Networks

Abstract

Hypergraph neural networks (HGNNs) have recently attracted much attention from researchers due to the powerful modeling ability. Existing HGNNs usually derive the data representation by capturing the high-order adjacent relations in a hypergraph. However, incomplete exploration and exploitation of hypergraph structure result in the deficiency of high-order relations among the samples. To this end, we propose a novel hypergraph convolutional networks (M-HGCN) to capture the latent structured properties in a hypergraph. Specifically, two novelty designs are proposed to enhance the expressive capability of HGNNs. (1) The CNN-like spatial graph convolution and self-adaptive hypergraph incidence matrix are employed to capture both the local and global structural properties in a hypergraph. (2) The dual-attention scheme is applied to hyperedges, which can model the interactions across multiple hyperedges to form hypergraph-aware features. The experimental results on the benchmark citation network datasets demonstrate the superior performance of the proposed method over the existing strong baselines.