Validation and Optimization of Link Prediction in Knowledge Graph Embeddings through Relationship Prediction

Common inference tasks in knowledge graphs include link prediction, relation prediction, and entity alignment. Knowledge graph embedding (KGE) has demonstrated its effectiveness for these tasks, with numerous KGE models achieving significant results in this domain. Nevertheless, given the intricate relation patterns in knowledge graphs, KGE models frequently show constrained reasoning capabilities, especially in link prediction. Notably, most KGE models have demonstrated satisfactory performance in relation prediction tasks. Motivated by this observation, this paper evaluates and analyzes relation prediction results for two common KGE models and proposes a novel inference method. Our method leverages relation prediction scores to support and optimize KGE models’ link prediction abilities. Comprehensive and rigorous experiments validate our methodology, achieving competitive results across both benchmark datasets.