Identifying scientific collaboration opportunities based on heterogeneous hypergraph link prediction

The rapid growth of scientific publications and patent records intensifies information overload and makes it increasingly difficult to identify promising collaboration opportunities across academic and technological domains. To address this challenge, we propose an event centric heterogeneous hypergraph framework that reframes collaboration discovery as a document conditioned, multi type recommendation problem. Instead of formulating collaboration discovery as a pairwise link prediction task, our method treats each paper or patent as a query event and constructs an event induced hyperedge that jointly connects its associated researchers, institutions, and research topics. This event centric formulation preserves collaboration events as holistic units and explicitly retains high order, multi entity context across publication and patent domains. Building on this representation, we develop a heterogeneous hypergraph variational autoencoder that jointly infers event level and entity level latent variables and reconstructs node hyperedge incidences. We further incorporate time aware entity activity profiles and type aware aggregation to capture temporal dynamics and heterogeneous event composition. With these components, the framework supports fine grained, document level recommendation by producing ranked lists of potential collaborators, institutions, and research topics conditioned on a specific paper or patent, thereby enabling multi perspective and structured collaboration planning in practical settings. Experiments and case studies on a real world cross domain semiconductor materials corpus demonstrate that the proposed formulation improves recommendation quality over strong baselines and helps uncover cross domain collaboration opportunities that single source approaches often miss.