About the project
One of the most exciting frameworks to quantify uncertainty of predictions is conformal prediction (CP). Under appropriate exchangeability conditions, it provides a confidence set (credibility set if prediction is Bayesian) for a multivariate estimate with statistical coverage guarantees. This PhD project will develop new CP methods for knowledge graphs (KGs), which are one of the most popular approaches for (semi-)structured data. There are many learning tasks that are based on KGs, such as KG completion, link prediction, and node and graph classification. Graph Neural Networks (GNNs) are very successful for learning on KGs and solving the mentioned tasks, but also have great potential for incorporating symbolic knowledge due to strong connections between GNNs and logics. We will develop methods for logic-aware CP on KGs using GNN for prediction and design new algorithms with theoretical guarantees. Then we will verify practical applicability and usefulness of these ideas and algorithms on benchmarks and on challenging real-world settings at BOSCH, with the possibility to develop industrial standard.
Project team
Professor Arnoldo Frigessi, University of Oslo
Associate Professor Egor Kostylev, University of Oslo, Department of Informatics
PhD Fellow Lorenzo Mastrandrea, University of Oslo, Department of Informatics