Link prediction, which aims to predict missing links or possible future links between two nodes, is one of the most important research in social network analysis. Higher-order link prediction, a natural extension of this problem in hypergraphs, focuses on predicting hyperlinks among multiple nodes. The extended problem has a wide range of applications, including predicting chemical reactions and forecasting social communications. Hypergraph neural networks (HGNNs), designed to address the characteristics of higher-order networks (hypergraphs), are variants of graph neural networks (GNNs). Compared with traditional methods, these end-to-end approaches based on HGNNs are effective tools for higher-order link prediction. However, these approaches still have some shortcomings. On the one hand, HGNNs, deriving from traditional GNNs, inevitably inherit unnecessary complexity and redundant computation from deep learning lineage. On the other hand, the existing aggregators do not take into account the similarity among the nodes' features, resulting in information loss in the embeddings of hyperlinks. To solve the both shortcomings, firstly, we propose a Light HyperGraph Neural Network (LHGNN) by removing some specific linear feature transformation layers and activation function layers to reduce complexity and increase reliability. Secondly, we propose a Hybrid Aggregator (HA) to obtain hyperlinks' embeddings more comprehensively by concatenating the mean and standard deviation of the nodes' embeddings. Experimental results on six datasets from four different domains show that our method outperforms state-of-the-art methods with a simpler structure.
