Scientific collaborator recommendation via hypergraph embedding

Abstract

Identifying potential scientific collaborators is critical to fostering innovation in an era of academic digitalization. Existing recommendation methods often rely on pairwise relations and fail to model the high-order, multi-relational nature of real-world collaboration networks. To address this, we propose a hypergraph embedding-based framework that constructs a heterogeneous Scientific Collaboration Hypergraph from the AMiner dataset. Using a hypergraph neural network and translational scoring, our method captures structural semantics and interdisciplinary patterns. The resulting graph contains 6,119 scholars, 18,092 publications, and nine types of hyperedges modeling diverse academic relations. Experimental results show that our approach achieves a Recall@10 of 0.1802, representing a 78% improvement over the strongest baseline. It also performs robustly in cold-start scenarios and generalizes well to interdisciplinary recommendations. A user study confirms the interpretability of the system, with Usefulness and Trust receiving average scores above 4.0 on a 5-point Likert scale. The proposed method demonstrates both effectiveness and transparency in collaborator recommendation.