Multi-hop Inference Networks Based on Contrastive Learning

In response to the existing issues of poor explainability and underperformance in sparse knowledge graphs for entity prediction, this paper introduces a Contrastive MultiHopNet (CMHN) approach based on contrastive learning to enhance the accuracy of tail entity prediction in multi-hop reasoning on knowledge graphs. The CMHN model incorporates an improved strategy for multi-hop path generation and selection, utilizing the AnyBURL rule mining method to generate multi-hop paths from head entities to tail entities as sequence-level positive samples. To obtain the crucial negative samples in contrastive learning, the CMHN model employs three negative sampling strategies (in-batch, previous-batch, and self-negative sampling), using these samples at both the sequence and entity levels for contrastive examples. Experiments demonstrate that the CMHN model outperforms comparative algorithms in terms of accuracy in tail entity prediction and the explainability of multi-hop inference paths.