Knowledge reasoning based on graph neural networks with multi-layer top-p message passing and sparse negative sampling

Graph Neural Networks (GNNs) have received increasing interest in knowledge reasoning since they can learn the structure and semantic information of graphs. However, as the number of hops increases, the number of entities grows exponentially, resulting in a sharp rise in resource costs. Furthermore, the presence of irrelevant entities during message passing can accumulate noise, thereby diminishing the model’s accuracy. To solve these problems, a knowledge reasoning model based on graph neural networks with multi-layer top-p message passing and sparse negative sampling is proposed. To be specific, we designed a dynamic top-p message-passing strategy that dynamically samples key entities related to the query based on their probability distribution, thereby reducing high resource costs. Then the similarity-based negative sampling is applied to dense entities in the knowledge graph, while random sampling is used for sparse entities, which enhances the model’s ability to identify irrelevant entities. Extensive experiments conducted on three datasets (WN18RR, FB15k-237, and NELL-995) demonstrate our model outperforms other SOTA methods in knowledge reasoning, achieving an average improvement of 5.16%, 6.16% and 6.71% in MRR, Hits@1 and Hits@10 respectively, and it also has a great advantage in training efficiency than SOTA GNN-based methods, with the average improvement of 11.99% in training time. Our model not only offers a novel method for knowledge reasoning, but also contributes valuable insights for the advancement of knowledge graph technologies.