Abstract: Long-tail relations containing only a small number of facts are prevalent in real-world knowledge graphs, and few-shot knowledge graph completion aims to address this data sparsity issue. However, existing approaches often neglect the inherent uncertainty of entities and triples, which limits model reasoning performance under noisy or data-scarce conditions. This paper proposed a covariance-optimized Gaussian metric learning model for uncertain completion (CoGMUC) model to tackle uncertain reasoning in few-shot knowledge graphs. This model represented entities and relations within knowledge graphs as Gaussian distributions, effectively capturing their inherent uncertainty through covariance matrices. It computed semantic similarity with a covariance-aware multi-matching network to complete missing facts and predict confidence levels. Furthermore, a difficult negative sample mining strategy was introduced to enhance the discriminative capability and generalization performance of the model. Experimental results on the public datasets NL27K and CN15K demonstrate that compared with the existing few-shot uncertain knowledge graph completion model based on Gaussian metric learning, CoGMUC improves mean reciprocal rank (MRR) by 21.8% and 2.3% and increases Hits@10 by 9.6% and 21.5%, respectively in the link prediction task. Meanwhile, in the confidence prediction task, the mean squared error (MSE) is reduced by 14.3% and 7.7%, respectively. The findings demonstrate that the CoGMUC model effectively models and leverages uncertainty information, significantly enhancing the performance of few-shot knowledge graph completion.

Key words: knowledge graph, uncertainty, knowledge graph reasoning, Gaussian metric learning, covariance modeling