Pretrained Node-Enhanced Graph Pruning Network for Inductive Link Prediction

Inductive link prediction for knowledge graph (KG) aims to reason missing relations between two unseen entities. Currently, the mainstream subgraph-based inductive link prediction frameworks learn the semantic logic within local subgraphs centered on target entities and initializes the nodes with uniform positional node features, ignoring the rich structural features within nodes, which leads to missing node information, low expressive ability, and weak discriminative properties, limiting the effectiveness of link prediction, thereby restricting the effectiveness of link prediction. To address the above limitations, we propose a model named Pretrained Node-Enhanced graph Pruning Network (PNEPN) for inductive link prediction by enhancing the structural information and relational semantics of the nodes attribute features while pruning the less relevant edges in the subgraph. Specifically, the proposed PENPE: (1) Employs a pretrained inductive graph neural network to learn node embedding to capture the structural feature and relational semantics of nodes. (2) Extracts the first-order structural feature of the nodes as initial features for the pretrained network to enrich the structural information of node attributes. (3) Introduces a subgraph pruning mechanism based on attention score to optimize the structure of subgraphs. Experimental results on dataset such as WN18RR, FB15k-237 and NELL-995 show that PNEPN outperforms existing methods on the inductive link prediction task.