SCHG: Spectral Clustering-guided Hypergraph Neural Networks for Multi-view Semi-supervised Learning

Multi-view semi-supervised learning enables to efficiently leverage multi-view information as well as labeled and unlabeled data to solve practical problems. With graph neural networks, multi-view semi-supervised learning can be smooth and robust to the label propagation process. Hypergraph learning is an approach to hypergraph topology that aims to identify and exploit high-order relations on hypergraphs to uncover data beyond one-to-one in real-world applications. However, traditional hypergraph construction methods usually consider only local correlations between samples and may ignore dependencies that exist in the wider context of the dataset. In this paper, we propose a novel multi-view high-order correlation modeling method, where the connectivity of hyperedges is determined through clustering, and complementary information from each view is integrated via a hypergraph neural network. Inspired by the divisibility of graphs revealed by spectral graph theory, the proposed method works well to capture global high-order correlations within data and uncover potential manifolds. To assess the effectiveness of hypergraph modeling, we conduct a comprehensive evaluation of a multi-view semi-supervised node classification task. The experiments illustrate that the proposed approach achieves superior performance compared to current state-of-the-art algorithms and general hypergraph learning across eight datasets.