HyIE: An internal-external induced embedding for knowledge hypergraph link prediction

Abstract

Knowledge hypergraphs have the widespread availability due to the ubiquity of $n$-ary relational facts in the real world. Link prediction over knowledge hypergraphs has emerged as a promising fundamental task in various domains, such as biology and social networks. However, existing approaches fail to consider the external information of $n$-ary tuples and extract the sequential information of entities within $n$-ary tuples, which leads to the performance bottleneck. To address this challenge, in this paper, we propose a novel knowledge hypergraph link prediction model, called HyIE. Specifically, by introducing virtual nodes, we design a hypergraph convolutional neural networks, called V-HGCN, to capture external structural information. To extract sequential information of entities within $n$-ary tuples, a relation-aware model equipped by Mamba tailored for knowledge hypergraphs is proposed, named HyMamba. Furthermore, to enhance the performance, we develop three negative sampling methods, namely, adversarial learning negative sampling, intra-loop negative sampling and degree-based negative sampling. Extensive experiments on real-world datasets have demonstrated that our HyIE outperforms the state-of-the-art models. Code for HyIE is available at https://github.com/nldmz/maincode.