Inductive Type-aware Reasoning over Knowledge Graphs

The primary objective of reasoning over Knowledge Graphs (KGs) is to derive novel facts based on existing ones. Inductive reasoning models have predominantly focused on predicting missing facts by acquiring logical rules. However, despite the significance of inductive relation prediction, most recent studies have been limited in a transductive framework, lacking the capability to handle previously unseen entities. Nonetheless, the subgraph mining methods often overlook the importance of entity type or the relational path, thereby limiting their comprehensive reasoning capabilities. To address these challenges, we propose a novel approach Inductive Type-awaRe LInk Prediction, called TRIP. In TRIP, we enhance the modeling of subgraph representations in a comprehensive manner, combining both latent type features and relational paths. Besides, we leverage mutual information and contrastive learning for knowledge graphs. Extensive experiments are conducted on two fully-inductive datasets, and TRIP outperforms baseline methods in terms of predictive accuracy and performance. It validates the effectiveness and usefulness of TRIP in exploring node neighboring relations on a global scale to characterize node features and reason over relational paths.