In the project we develop new techniques and tools for AI tasks where the interaction of two or more objects is to be predicted. Such problems are prevalent in society, found in numerous applications, such as various link prediction problems in social networks, recommendation systems, among others. In the biomedical domain, prediction of binding of drug molecules to their molecular targets and responses of a patient to a combination of drugs are crucial interaction prediction tasks for enabling personalized and precision medicine.

The project advances state of the art in AI by tackling several key challenges related to interaction prediction with tensor methods: (i) learning new representations for structured objects to increase accuracy and explainability, (ii) developing methods for partially-observed fragmented datasets, (iii) increasing explainability of predictions through integrating symbolic prior knowledge in learning, (iv) data-efficient learning of models through active learning and structured budgets, as well as (v) new methods for trustworthy performance evaluation especially in cold start scenarios, where interactions of previously unseen objects need to be predicted.