Page 23 - Sirius_Annual_Report_2021

exists (e. g., in the Wikidata knowledge graph) that can be used to contextualize the data so that the performance
of the ML approach could be improved. Thus, we arrive at research questions such as: Given tabular data, how to find openly available domain knowledge in the Web of Data that could contextualize given tabular data? How to align tabular data to entities and properties in an RDF dataset (i. e., the domain knowledge represented in RDF format, which is the common format in the Semantic Web)? How to find
out which parts of the external knowledge help most,
so that when this information is improved it will have a significant impact on the performance? Which parts of the data have a negative impact on the performance and should thus be removed? How exactly can external knowledge be incorporated into an ML approach such as in a preprocessing step to improve the quality of existing training data? How can domain knowledge help in post- processing the output of the ML approach? How can the solution space exploration of the ML approach be guided by domain knowledge? How can the search space be pruned by making use of domain knowledge?
The situations and research questions that we collect can be ordered in two groups: those research questions that primarily address domain adaptation, such as, how to make use of (medical) taxonomies while training a (disease) classification model? Secondary research questions do not directly address domain adaptation but enable domain adaptation, e. g., em- bedding of knowledge graphs into vector spaces so that these can be processed by classical ML approaches, or they tackle the improvement of domain knowledge so that the perfor- mance of a domain adapted approach can be improved further, e. g., by anomaly detection in knowledge graphs and automatic knowledge graph completion.
The task of identifying and describing situations is both a top-down approach (guided by combinatorial exploration and brainstorming, zooming out and in) and a bottom-up approach (inspired by existing research). Not only does it allow us to structure existing approaches according to situations,
but it also lets us identify research questions that have so far not been addressed sufficiently. Our vision is that we can develop a methodology for domain adaptation, where an individual or organization browses the graph of situations to learn about how to realize domain adaptation or at least finds pointers to relevant sources such as publications.
Knowing what research could be done, which is an outcome of our activities, needs to be complemented by what is relevant for our SIRIUS partners, so that we can focus on those tasks or challenges especially relevant to our partners for mutual benefit. Therefore, we plan a stronger involvement of our partners in 2022 to prioritize our activities and to create a roadmap for research.
Beyond domain adaptation or improving domain knowledge that can then be used for domain adaptation, we actively work on a couple of other topics, within a group that also includes non-SIRIUS researchers at the UiO Department
of Informatics (i. e., Anne-Marie George, Thomas Kleine Büning, and Meirav Segal). For example, we investigate in which way domain knowledge can assist reinforcement learning tasks. UiO researchers in the NRC project «Safe and Beneficial Artificial Intelligence» investigate active learning problems involving human and societal preferences such
as learning preferences from interactions, collaborating effectively with humans, and making repeated decisions that are fair in the long term. Here, background knowledge, e. g., behavioral conventions, structure of preferences, features, and their relations of states of the world, might be able to improve the learning. Reversely, elicitation schemes might be able to gather and structure knowledge.
Furthermore, we carry out research related to explainable AI (XAI). The current approaches and technologies in XAI mostly focus on shedding light on the behavior of black-box machine learning models (like deep neural networks) by explaining their decisions to the users. However, the least work has been done towards employing the information provided by the explanations for enhancing the models concerning accuracy, fairness, and robustness in a systematic way. In SIRIUS, we have studied this research area and devised explanation-based frameworks for investigating the accuracy and robustness of black-box ML classification models [1].
Figure 1. Image taken from [2]. Fine tuning model architecture where each component is shown with inputs and outputs. tc and ts are knowledge graph triples relating to chemicals and species each with a score SF and loss l. c, s, and are the prediction input variables while ŷ is the predicted toxicity.
SIRIUS ANNUAL REPORT 2021
| 23