HOPLoP: multi-hop link prediction over knowledge graph embeddings

Large-scale Knowledge Graphs (KGs) support applications such as Web search and personal assistants and provide training data for numerous Natural Language Processing tasks. Nevertheless, building KGs with high accuracy and domain coverage remains difficult, and neither manual nor automatic efforts are up to par. Link Prediction (LP) is one of many tasks aimed at addressing this problem. Its goal is to find missing links between entities in the KG based on structural by exploiting regularities in the graph structure. Recent years have seen two approaches emerge: using KG embeddings, and modelling complex relations by exploiting correlations between individual links and longer paths connecting the same pair of entities. For the latter, state-of-the-art methods traverse the KG itself and are hampered both by incompleteness and skewed degree distributions found in most KGs, resulting in some entities being overly represented in the training set leading to poor generalization. We present HOPLoP: an efficient and effective multi-hop LP meta method that performs the equivalent to path traversals on the KG embedding space instead of the KG itself, marrying both ideas. We show how to train and tune our method with different underlying KG embeddings, and report on experiments on many benchmarks, showing both that HOPLoP improves each LP method on its own and that it consistently outperforms the previous state-of-the-art by a good margin. Finally, we describe a way to interpret paths generated by HOPLoP when used with TransE.