TransE, from the paper “Translating Embeddings for Modeling Multi-relational Data” presented at NIPS 2013 by Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko, is one of the foundational methods for embedding knowledge graphs. A knowledge graph stores facts as triples of the form (head entity, relation, tail entity), such as (Paris, capital-of, France), and TransE learns to place entities and relations in a shared vector space.
The model’s defining idea is elegantly simple: a relation is represented as a translation, a single vector that you add. If a triple is true, then the head entity’s vector plus the relation’s vector should land very close to the tail entity’s vector. Training adjusts all the vectors so that true triples satisfy this approximately and false ones do not. This geometric picture, where relations are arrows you follow, is easy to interpret and cheap to compute.
That simplicity was the point. Earlier methods for multi-relational data used far more parameters and struggled to scale. TransE achieved strong link-prediction performance, predicting missing facts in a knowledge base, while remaining efficient enough to handle datasets with millions of entities and relations.
For businesses building or using knowledge graphs, such as search engines, recommendation systems, and enterprise data integration, TransE showed that the messy web of facts could be turned into clean numerical vectors usable by standard machine learning, kicking off a large family of knowledge graph embedding techniques.