Missing-edge aware knowledge graph inductive inference through dual graph learning and traversing

Knowledge graph (KG) is a kind of structured human knowledge of modeling the relations between real-world entities. This paper studies the KG inductive inference problem, i.e., predicting the relations for out-of-KG entities. However, due to the incomplete nature of the KGs, the connections of some relations are missing. This makes existing differentiable rule learning methods unable to represent some possible rule candidates, which will further affect the inductive inference result. To solve this challenge, our research hypothesis is that the semantics of relation's argument can be well used to reflect the possible connections between relations. We propose a KG inductive inference model, RuleNet, which consists of two parts. Firstly, a query-dependent dual graph construction method is proposed, which is able to learn the relation connections using the information of the relation's argument. Secondly, a dual graph traversing method is proposed, which is able to traverse all possible rule candidates even if some rules cannot be formed due to the missing edges. Performance of the proposed methods is evaluated using the FB15K237 (10%–20%), WN18RR (10%–20%) and YAGO3-10 (10%–20%) benchmarks. Experimental results show that RuleNet achieves a superior performance compared with many strong baselines. Ablation studies have verified the effectiveness of the proposed network components. Qualitative analysis shows that RuleNet can learn meaningful dual graph and logic rules.