An Application-Oriented Taxonomy of Uncertainty Quantification Methods in Machine Learning

Uncertainty quantification (UQ) is essential for using machine learning (ML) models in high-stakes domains, yet the wide variety of available methods makes selecting an appropriate approach challenging. This paper introduces a literature-based taxonomy of UQ methods in ML from an application-oriented perspective. Instead of focusing on individual algorithms, the taxonomy groups techniques into the following method families: Bayesian, ensemble, likelihood-based, evidential and higher-order distribution, distribution-free, generative, density-based and calibration methods. Each family is described and compared using key selection criteria, namely the addressed source of uncertainty, integration depth, degree of model agnosticism, general performance, capability for detecting out-of-distribution (OOD) data, computational cost, integration effort, interpretability and explainability, and output type. The resulting taxonomy, summarized in a structured comparison table, provides practitioners with an application-oriented overview that supports informed method selection.