Knowledge graphs are large networks of facts connecting entities such as people, places, and organizations. A simple query like “Where was Beth Hart born?” can be answered by finding the entity Beth Hart in the graph and following the relation that connects her to her birthplace. A complex query consists of multiple conditions combined together, for example: “Which musicians have collaborated with Joe Bonamassa AND performed at the Royal Albert Hall?” Answering it means combining several simple facts (collaborations, venues performed at) into one query.

Knowledge graphs are usually incomplete: they’re built by collecting facts from limited sources, so some true facts (say, an obscure collaboration or a concert that was never logged anywhere) are simply missing. Because of this, my work is on a family of methods known as Complex Query Answering (CQA) that still find good answers even when facts are missing, by ranking candidate entities by how likely they are to be correct. The problem is that most CQA methods work like a black box, giving a ranking without explaining why. My research focuses on opening this black box: we introduced the first explainability approach in this domain, CQD-SHAP, which measures how much each part of a query contributes to an answer’s ranking, and we’re now exploring approaches that work no matter which model produced the ranking, as well as other ways to explain these answers.

If any of this sounds interesting to you and you’d like to collaborate, I’d love to hear from you!

Data Science Junior Research Group
Fürstenallee 11, 33102 Paderborn, Germany
Room FU.201.1

View larger map