Evidence–Theoretical Modeling of Uncertainty Induced by Posterior Probability Distributions

We discuss how the posterior probability distributions produced by machine learning models for analyzed objects can be transformed into evidence-theoretical mass functions that model uncertainties associated with operating those distributions. We investigate the mathematical properties of the introduced mass functions and their corresponding belief functions. We also construct some uncertainty measures based on the functions considered and compare them with several classical uncertainty measures, both theoretically and practically, in the active learning scenarios.