Hyperbolic Embedding of Subsumption Groups in Large-Scale Biomedical Ontologies with Containment Cone Constraints

Embedding methods are a key enabler of recent advances in language model and augmented retrieval systems. For biomedical ontologies in particular, embedding data in a geometric space with non-zero curvature can capture the structure of hierarchies more robustly. In this work, a hyperbolic embedding method is proposed which is equipped with containment constraints for modeling subsumption groups. On synthetically generated directed graphs, it is found that arbitrarily low distortions can be achieved. Through experimentation on a real-world biomedical ontology, it is shown that our method satisfies a larger portion of containment properties compared to existing models.