Abstract
Hyperbolic space has proven to be well-suited for capturing hierarchicalrelations in data, such as trees and directed acyclic graphs. Prior workintroduced the concept of entailment cones, which uses partial orders definedby nested cones in the Poincar\'e ball to model hierarchies. Here, we introducethe ``shadow cones" framework, a physics-inspired entailment cone construction.Specifically, we model partial orders as subset relations between shadowsformed by a light source and opaque objects in hyperbolic space. The shadowcones framework generalizes entailment cones to a broad class of formulationsand hyperbolic space models beyond the Poincar\'e ball. This results in clearadvantages over existing constructions: for example, shadow cones possessbetter optimization properties over constructions limited to the Poincar\'eball. Our experiments on datasets of various sizes and hierarchical structuresshow that shadow cones consistently and significantly outperform existingentailment cone constructions. These results indicate that shadow cones are aneffective way to model partial orders in hyperbolic space, offering physicallyintuitive and novel insights about the nature of such structures.