Knowledge graphs capture domain knowledge in the form of facts about entities and have many use cases in information systems. One of these is link prediction which is the task of inferring new facts in a knowledge graph. Even though research in the field gains momentum, few models focus on being able to explain their predictions. In this paper, we introduce ResqLP, a link prediction model that learns probabilistic first-order logic rules from both entity semantics and sequences of relations. The model includes a feature extraction phase that, by the use of enclosing subgraphs, aims to cope with the inherently hard path enumeration problem. We also present a general way to capture entity semantics based on relations found in entity neighbourhoods.
We find that using enclosing subgraphs speeds up feature extraction even though large knowledge graphs still cause problems in some cases. Furthermore, even though the quality varies, ResqLP is able to learn non-trival logic rules. The model performs the best on relatively sparsely connected knowledge graphs with many different relations.