Adaptive Relation-Specific Ensemble Learning for Knowledge Graph Completion

Abstract

Knowledge graph embedding (KGE) models learn representations of entities and relations to infer missing links in knowledge graphs. Existing KGE models employ different scoring functions to capture diverse relational patterns, yet no single model effectively represents all relation types. Consequently, ensemble methods have been proposed to combine multiple pretrained KGE models. However, existing ensembles are relation-independent, overlooking that different KGE models excel on different relation types. In this work, we propose Adaptive Relation-specific Weighted Ensemble (AReWE), a relation-aware ensemble framework that jointly fine-tunes pretrained KGE models.1 AReWE combines global optimization through a novel relation-specific batching strategy with local specialization via adaptive relation-specific weighted losses. Experiments on FB15K237, WN18RR, and NELL-995 demonstrate consistent improvements over existing ensemble methods and achieve state-of-the-art performance on FB15K237 and NELL-995.