Semi-supervised Learning with Graph Convolutional Networks Based on Hypergraph

Graph convolutional networks (GCNs), which rely on graph structures to aggregate information of neighbors to output robust node embeddings, have been becoming a popular model for semi-supervised classification tasks. However, most existing GCNs ignore the importance of the quality of graph structures, therefore output suboptimal classification performance. In this paper, we propose a new graph learning method to output a high-quality graph structure, aiming at eventually improving classification performance for the downstream GCN model (HS-GCN). Specifically, the proposed graph learning method employs an adaptive graph learning to capture the intrinsic low-level correlation of data, and learns the more useful high-level correlation from a hypergraph. Besides, sparse learning and a low-rank constraint are integrated with graph learning respectively to remove redundant information, and to obtain a compact graph structure for promoting information aggregation of GCNs. The experimental results show that the graph structure of our proposed graph learning method can significantly improve the classification performance of GCNs.