Elliptical Ordinal Embedding

Ordinal embedding aims at finding a low dimensional representation of objects from set constraints the form ”item j is closer to item i than k”. Typically, each object mapped onto point vector in metric space. We argue that mapping density instead provides some interesting advantages, including an inherent reflection uncertainty about itself and its relative location Indeed, this paper, we propose embed as Gaussian distribution. investigate ability these embeddings capture underlying structure data while satisfying constraints, explore properties representation. Experiments on synthetic real-world datasets showcase advantages our approach. In addition, illustrate merit modelling uncertainty, which enriches visual perception