Knowledge graph completion with mixed geometry tensor factorization

Аннотация

In this paper, we propose a new geometricapproach for knowledge graph completionvia low rank tensor approximation. We augmenta pretrained and well-established Euclideanmodel based on a Tucker tensor decompositionwith a novel hyperbolic interactionterm. This correction enables more nuancedcapturing of distributional propertiesin data better aligned with real-world knowledgegraphs. By combining two geometriestogether, our approach improves expressivityof the resulting model achieving new stateof-the-art link prediction accuracy with a significantlylower number of parameters comparedto the previous Euclidean and hyperbolicmodels.