Hypergraph Convolutional Network with Hybrid Higher-Order Neighbors.

Hypergraph-based methods can learn non-pairwise associations more efficiently in many real-world datasets. However, existing hypergraph-based methods do not consider the relationship of the hybrid neighborhood. To address this issue, we propose a hybrid higher-order neighborhood based hypergraph convolutional network (HybridHGCN). Technically, feature embeddings are generated via k-hop hypergraph convolution layers and mixed by the hybrid message operator. To evaluate the proposed HybridHGCN, we conduct experiments on the citation network datasets and the visual object datasets. The experimental results show that HybridHGCN brings significant improvements over state-of-the-art hypergraph neural network baselines.