Abstract
Geometric aspects of knowledge graph embedding models directly impact their capability to preserve knowledge from the original graph to the vector space. For example, the capability to preserve structural patterns such as hierarchies, loops, and paths present as relational structures in a knowledge graph depends on the underlying geometry. In these years, temporal information has gained lots of attention from researchers. While non-Euclidean geometry, e.g. Hyperbolic Geometry, has been shown to work well in static knowledge graph embedding models for such relational structures, this does not hold for temporal information in knowledge graphs. This is due to the different characteristics of temporal information: time can be seen mostly as a linear construct and using a geometry that is not suitable for this can adversely affect performance. To address this research gap, we provide a novel temporal knowledge graph embedding model that combines different geometries: the non-temporal part of the knowledge is mapped to a hyperbolic space and the temporal part is mapped to a Euclidean space. Our extensive evaluations on several benchmark datasets show a significant performance improvement in comparison to state-of-the-art models.
Users
Please
log in to take part in the discussion (add own reviews or comments).
