HySAE: An efficient semantic-enhanced representation learning model for knowledge hypergraph link prediction

Author Identifier

Jianxin Li: https://orcid.org/0000-0002-9059-330X

Document Type

Conference Proceeding

Publication Title

WWW 2025 - Proceedings of the ACM Web Conference

First Page

86

Last Page

97

Publisher

Association for Computing Machinery, Inc

School

School of Business and Law

RAS ID

79347

Funders

National Natural Science Foundation of China (62472311) / Key Research and Development Program of Ningxia Hui Autonomous Region (2023BEG02067)

Comments

Li, Z., Wang, X., Zhao, J., Feng, F., Chen, Z., & Li, J. (2025). HySAE: An efficient semantic-enhanced representation learning model for knowledge hypergraph link prediction. Proceedings of the ACM on Web Conference 2025, 86-97. https://doi.org/10.1145/3696410.3714549

Abstract

Representation learning technique is an effective link prediction paradigm to alleviate the incompleteness of knowledge hypergraphs. However, the n-ary complex semantic information inherent in knowledge hypergraphs causes existing methods to face the dual limitations of weak effectiveness and low efficiency. In this paper, we propose a novel knowledge hypergraph representation learning model, HySAE, which can achieve a satisfactory trade-off between effectiveness and efficiency. Concretely, HySAE builds an efficient semantic-enhanced 3D scalable end-to-end embedding architecture to sufficiently capture knowledge hypergraph n-ary complex semantic information with fewer parameters, which can significantly reduce the computational cost of the model. In particular, we also design an efficient position-aware entity role semantic embedding way and two enhanced semantic learning strategies to further improve the effectiveness and scalability of our proposed method. Extensive experimental results on all datasets demonstrate that HySAE consistently outperforms state-of-the-art baselines, with an average improvement of 9.15%, a maximum improvement of 39.44%, an average 10.39x faster, and 75.79% fewer parameters.

DOI

10.1145/3696410.3714549

Access Rights

subscription content

Share

 
COinS
 

Link to publisher version (DOI)

10.1145/3696410.3714549