Using hyperboloids to embed and query knowledge graphs

What is it about?

Hyperboloids are bounded curved surfaces — curved analogues of rectangles — and characteristics of hyperbolic space enable them to capture hierarchical information that conventional knowledge graph embeddings lose. Our paper describes a neural-network architecture for learning hyperboloid embeddings that enables us to logically compose knowledge graph queries. We could, for instance, search a product graph for all footwear from both brand A and brand B, a query that could be logically represented as the union of the intersections of the embeddings for brand A and brand B with the embedding for footwear.

Featured Image

Photo by Markus Winkler on Unsplash

Photo by Markus Winkler on Unsplash

Why is it important?

Knowledge graphs are an extremely efficient way to represent information, but the standard techniques for analyzing them, which involve tracing connections through a graph one hop at a time, don’t scale well. Recently, knowledge graph embeddings, which represent elements of the graph as points in a multidimensional space, have provided a way to analyze knowledge graphs more efficiently. But they give up much of the graphs’ informational richness.

Perspectives