HTransE: Hybrid Translation-based Embedding for Knowledge Graphs

Abstract

Basically, a Knowledge Graph (KG) is a graph variant that represents data via triplets comprising a head, a tail, and a relation. Realistically, most KGs are compiled either manually or semi-automatically, and this usually results in a significant loss of vital information with respect to the KG. Thus, this problem of incompleteness is common to virtually all KGs; and it is formally defined as Knowledge Graph Completion (KGC) problem. In this paper, we have explored learning the representations of a KGs with regard to its entities and relations for the purpose of any predicting missing link(s). In that regard, this paper proposes a hybrid variant, composed of TransE and SimplE models, for solving KGC problems. On one hand, the TransE model depicts a relation as the translation from the source entity (head) to the target entity (tail) within an embedding space. In TransE, the head and tail entities are derived from the same embedding-generation class, which results in a low prediction score. Also, the TransE model is not able to capture symmetric relationships as well as one-to-many relationships. On the other hand, the SimplE model is based on Canonical Polyadic (CP) decomposition. SimplE enhances CP via the addition of the inverse relation, while the head entity and tail entity are derived from different embedding-generation classes which are interdependent. Hence, we employed the principle of inverse-relation embedding (from the SimplE model) onto the native TransE model so as to yield a new hybrid resultant: HTransE. Therefore, HTransE boasts of efficiency as well as improved prediction scores. Efficiently, HTransE converges much quicker in comparison to TransE. In other words, HTransE converges at approximately <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$n/2$</tex> iterations where <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$n$</tex> denotes the iterations required to fully train TransE. Our results outperform the native TransE approach with a significant difference. Also, HTransE outperforms several state-of-the-art models on different datasets.