Entity neighborhood awareness and hierarchical message aggregation for inductive relation prediction

Inductive relation prediction aims to apply the reasoning ability learned from existing knowledge graphs to predict the relation between invisible entities. Recently proposed subgraph-based models achieve inductive reasoning by encoding the enclosing subgraph for target triplet. Such models overlook the non-public neighborhood information of the target entity pair. Moreover, modeling the relation paths that compose the logical rules is a common challenge in subgraph reasoning. To address these issues, we propose a novel subgraph-based model named ENarMa, for inductive reasoning. Firstly, an entity neighborhood awareness method is introduced to calculate the local subgraph for each target entity, thereby obtaining the non-public neighborhood information for the target triplet. Then, we propose a hierarchical message aggregation mechanism, where the node layer encodes representations for different nodes, and the path layer integrates the combined attention to model relation path features. This mechanism can further highlight important local evidence while encoding graph semantics. Extensive experiments conducted on WN18RR, FB15K-237 and NELL-995 demonstrate the superiority of ENarMa. Specifically, the average performance of ENarMa exceeds the suboptimal value by an average score of 80.61 in Hits@10 and 66.28 in MRR.