1. Project Title: Semantic explanations of predictions using knowledge graphs
2. Area of knowledge: Physical Sciences, Mathematics and Engineering Panel
3. Group of disciplines: Theoretical and Applied Mathematics, Computer Sciences and IT
4. Research project:
The increasing pervasiveness of AI in every sphere of human endeavour has brought about the need to help users of AI-based systems understand their outcomes.
Recent work has successfully produced explanations for complex models based on assumptions like linearity in the locality of each instance, while other approaches are based on similarities between predictions and training set data points. However, to provide users with informative explanations of predictions we need to fulfil the properties of human understanding, namely, that human explanations imply social interaction which is grounded in a shared context, and that users select explanations from a large space of possible explanations based on their own understanding of the context. Existing approaches typically focus on data-based explanations and thus lack the semantic context needed to fulfil the requirements for human-centric explanations.
This project proposes to develop novel approaches for explainable AI that employ ontologies and knowledge graphs to provide the semantic context for machine learning prediction explanations. When data is described formally with ontology concepts and linked into knowledge graphs, we can have a shared understanding of its meaning. This supports building informative explanations, as well computing semantic similarities between relevant examples.
The life sciences and healthcare domains represent unparalleled opportunities for the successful application of the proposed approaches. They are data-rich, raw data is often semantically meaningful, there are more than 2 billion biomedical entities in knowledge graphs, and lastly, as black-box models, such as deep learning, are increasingly adopted due to their increased performance, the need to provide informative explanations to users grows.
5. Job position description:
Candidates should have an MSc or equivalent degree in Computer Science, Bioinformatics, Mathematics, Statistics, Life Sciences or related areas. Candidates should be proficient in programming and the English language. Multidisciplinary backgrounds are welcome. Research experience is a plus, particularly in machine learning and/or semantic web.
The successful candidate will conduct research to address the following challenges: (1) developing methods to automate the selection of ontologies/knowledge graphs that take into account the personalization of explanations; (2) designing improved methods for mapping input raw data to ontologies, to ensure explanations can be created even when data has no semantic layer; (3) investigating and improving different approaches for the selection of representative data points; (4) designing methods for building explanations from knowledge graphs that take into account cognitive load and semantic coverage. The evaluation of explanation performance will be achieved through functional, human and application grounded evaluation approaches, with a strong emphasis on biomedical and clinical data.
The work will be supervised by Catia Pesquita (https://scholar.google.com/citations?user=8nGYegMAAAAJ), Assistant Professor at the Department of Informatics, Faculty of Sciences (FCUL), University of Lisbon. She is also senior Researcher at LASIGE, where she coordinates the Health and Biomedical Informatics Research Line of Excellence. The candidate will integrate a multidisciplinary research team within LASIGE with longstanding expertise in knowledge graphs and semantic web technologies, and their application to data integration and mining in the biomedical domain.
This work plan is aligned with a collaboration with researchers from NEC Laboratories who are experts in developing explanations for predictions based on graph-based data. During the PhD, the student will have opportunities for visiting NEC Labs in Germany.
6. Group leader:
Full name: Catia Luisa Santana Calisto Pesquita
Email: clpesquita@ciencias.ulisboa.pt
Google scholar profile: https://scholar.google.com/citations?user=8nGYegMAAAAJ