An Application-Oriented Process Model for Selecting Uncertainty Quantification Methods in Machine Learning

This paper presents an application-oriented process model for selecting uncertainty quantification methods in machine learning. The work is motivated by the increasing relevance of uncertainty quantification in product engineering applications, including the automotive domain, where reliable decision making under uncertainty is essential. In this context, machine learning is widely used in data mining to derive knowledge about systems and user behavior from data and to make this knowledge accessible to product developers. As these applications increasingly influence engineering decisions, the assessment and communication of uncertainty become critical requirements. Each data mining task is unique and no single algorithm is optimal for all tasks. This observation motivates the need for a structured process model to support the selection of appropriate uncertainty quantification methods in machine learning. The proposed process model builds on established problem-solving methods and algorithm-selection support approaches as its conceptual foundation. On this basis, it translates the technical characteristics and assumptions of uncertainty quantification algorithms into a structured and systematic decision guide. The model does not focus on individual methods, but instead supports informed selection decisions across a broad and heterogeneous method space. In addition to technical aspects, the model incorporates practical barriers and priorities derived from ethnographic observations in research settings, capturing typical implementation and integration limitations. By accounting for these real-world considerations, the model reflects the conditions under which uncertainty quantification methods are typically applied in practice. The resulting selection process is transparent, reproducible, and domain agnostic. It systematically combines method diversity, technical boundary conditions, and real-world project constraints to support the targeted selection of suitable uncertainty quantification approaches. The contribution is made at a meta level, enabling structured navigation of the method space rather than providing a direct comparison of individual uncertainty quantification algorithms.