Modeling and Addressing Uncertainty in Artificial Intelligence - Contributions to Uncertainty-Aware Machine Learning and Explainable AI

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The rapid and widespread adoption of artificial intelligence (AI) presents significant opportunities for organizations and society. However, it also raises fundamental questions about the validity and transparency of AI-based decisions. Uncertainty is a central challenge in this context because it can originate from different sources and affect data, machine learning (ML) models, and their explanations. This dissertation investigates how to model and address uncertainty to support the valid and transparent use of AI in organizational and societal decision-making. It contributes theoretical foundations and methods for uncertainty-aware ML and explainable artificial intelligence (XAI). Regarding data uncertainty, a taxonomy is developed to systematically classify methods for addressing data quality defects in ML. Building on this taxonomy, the dissertation establishes theoretical foundations for data quality to provide a rigorous basis for understanding data-related phenomena and data uncertainty. Subsequently, two methods demonstrate how to address such uncertainty: one method propagates data uncertainty exactly through neural networks, and the other uses a language model to assess and improve the currency of textual data. Regarding model and explanation uncertainty, the dissertation examines how uncertainty influences the correctness of explanations for ML predictions. Sources of uncertainty affecting explanations are identified and conceptually distinguished to provide a basis for understanding and assessing explanation correctness. Furthermore, a method is developed for the exact reconstruction of ML model decision boundaries. By directly capturing the decision-making behavior of ML models, this method provides faithful and robust explanations while avoiding the uncertainty that arises from approximating the underlying decision boundaries.