An original model for multi-target learning of logical rules for knowledge graph reasoning

Abstract

Large-scale knowledge graphs are crucial for structuring human knowledge; however, they often remain incomplete. This paper tackles the challenge of completing missing factual triples in knowledge graphs using through rule reasoning. Current rule learning methods tend to allocate a significant portion of triples to constructing the graph during training, while neglecting multi-target reasoning scenarios. Furthermore, these methods typically depend on qualitative assessments of mined rules, lacking a quantitative method to evaluate rule quality. We propose a model that optimizes training data usage and supports multi-target reasoning. To overcome limitations in evaluating model performance and rule quality, we propose two novel metrics. Experimental results show that our model outperforms baseline methods on five benchmark datasets, validating the effectiveness of these metrics.