The paper discusses the limitations of the standard evaluation protocol for Knowledge Graph Completion methods, which involves ranking every entity of a Knowledge Graph. The authors argue that this task becomes too heavy for larger scale Knowledge Graphs and that random sampling of entities, used to mitigate this problem, does not produce accurate ranking metrics. They propose a new framework that uses relational recommenders to guide the selection of candidates for evaluation. They provide both theoretical and empirical justification for this methodology, demonstrating that it can reduce time and computational needs while improving the estimation accuracy.
Publication date: 2 Feb 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2402.00053
