Abstract

Knowledge graphs are large graph-structured databases of facts, whichtypically suffer from incompleteness. Link prediction is the task of inferringmissing relations (links) between entities (nodes) in a knowledge graph. Weapproach this task using a hypernetwork architecture to generate convolutionallayer filters specific to each relation and apply those filters to the subjectentity embeddings. This architecture enables a trade-off between non-linearexpressiveness and the number of parameters to learn. Our model simplifies theentity and relation embedding interactions introduced by the predecessorconvolutional model, while outperforming all previous approaches to linkprediction across all standard link prediction datasets.

 

Quick Read (beta)