A probabilistic ensemble approach for knowledge graph embedding

Abstract

Knowledge graph embedding (KGE) is a technique for embedding entities and relations of knowledge graphs (KGs) into continuous vector spaces while maintaining the inherent structure of KGs, in this way link prediction can be facilitated by scoring candidate triples. Existing KGE models only capture specific features of a KG, however, link prediction capacity heavily relies on comprehensive features of a KG. In this work, we propose an ensemble approach for KGE by incorporating the features captured by different KGE models. As a key step, we propose a probabilistic scoring index for characterizing the link prediction performance of a given KGE model. Based on this we established a theoretical framework to calculate the optimal parameters for the ensemble model and predict its corresponding link prediction rate. In particular, we proved that our ensemble approach can always improve link prediction performance under some assumptions. The Upper Confidence Bound Algorithm is then used to adjust the parameters. Experimental results on two widely used KGs show that the proposed ensemble approach achieves the state-of-the-art link prediction rate.