Explaining Link Predictions in Knowledge Graph Embedding Models

Session track: AI Engineering

Session time: 04:00 pm - 04:30 pm UTC November 12

Session description:

Knowledge Graph Embedding (KGE) models such as TransE, ComplEx, and DistMult are widely used for link prediction on knowledge graphs, but they lack direct interpretability; a growing concern as regulatory frameworks like the EU AI Act increase explainability requirements for AI systems. In this talk, we show how to explain KGE link predictions in practice. We walk through concrete examples: given a prediction such as gene PRKN is associated with Parkinson's Disease, how do we identify which training triples in the knowledge graph were most responsible for that score? We demonstrate this using real datasets (FB15k-237, WN18RR) and biomedical knowledge graphs from applications in gene-disease association and patient relapse risk prediction. We survey post-hoc explainability methods, including Gradient Rollback, Adversarial Explanations, ExamplE, and Kelpie, focusing on their practical tradeoffs: speed, memory, and whether the explanations actually make sense to an end user. We show code using AmpliGraph and discuss how to evaluate explanations using Remove-and-Retrain protocols.