Reproduce “The Role of Graph Topology in the Performance of Biomedical Knowledge Graph Completion Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alberto Cattaneo, Stephen Bonner, Thomas Martynec, Edward Morrissey, Carlo Luschi, Ian P Barrett, Daniel Justus
Researchers from Graphcore Research and AstraZeneca investigated how local topological properties within biomedical knowledge graphs affect the predictive accuracy of Knowledge Graph Embedding models. Their fine-grained, triple-level analysis revealed that entity degrees and compositional patterns correlate with model performance, while showing that simply expanding a knowledge graph with more data can, counter-intuitively, degrade performance for shallow embedding models.
Key insights
- Aggregate topological analyses (dataset or relation-type level) introduce noise and fail to establish clear links to KGE model accuracy; a fine-grained, triple-level analysis is essential.
- Entity degrees strongly correlate with KGE model performance: high tail in-degree for the same relation type is beneficial, while high head out-degree is detrimental.
- Compositional patterns consistently improve model accuracy, particularly for entities with small degrees, while the impact of symmetry/inverse patterns depends on their counterpart edges being present in training data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.