A Relation-Specific Entropy-Based Ensemble Approach for Knowledge Graph Embedding

Abstract

Knowledge Graph Embedding (KGE) aims to represent entities and relationships from knowledge graphs (KGs) in vector spaces. Existing KGE methods often focus narrowly on specific relation patterns or employ ensemble weights based on the prediction scores of KGE models, which constrain their link prediction capacity for KG with diverse relation patterns. In response to this limitation, we introduce a relation-specific entropy-based KG ensemble approach, denoted as ERSE. We assume that the entropy of the similarity distribution of prediction vectors is a reliable indicator of the inference potential for particular relations. On the basis of this hypothesis, ERSE utilizes relation-specific ensemble weights, determined by the entropy of the similarity distribution of prediction vectors derived from the base-model and training set. Moreover, ERSE is designed to be flexible in that it combines base-models using normalized entropy measurements in order to be effectively applied to various KG data sets and complex prediction environments. Using four translation- based knowledge graph embedding methods (TransE, TransH, TransR, and TransD) as base-models, our experiments on the FB15K and FB15k237 datasets show that ERSE surpasses both of each single-model and conventional ensemble approaches in terms of prediction accuracy. These results corroborate our hypothesis that the relation-specific entropy of prediction vector similarity can improve the inference performance of KGE.