min_expected_total_variation

probly.quantification.min_expected_total_variation = <flextype.singledispatch.Flexdispatch object>

Compute the distance-based epistemic uncertainty of a second-order distribution.

Defined by [SBCHullermeier24] as the smallest expected total-variation distance from Q to a point on the simplex:

\[EU(Q) = \tfrac{1}{2} \min_{q \in \Delta} \mathbb{E}_{p \sim Q}[\lVert p - q \rVert_1],\]

bounded by (K - 1) / K for K classes. Computed directly for a sample. For a parametric distribution such as a Dirichlet it is estimated by Monte-Carlo, and that implementation also takes num_samples and a generator.

Parameters:

distribution – A sample of categorical distributions or a Dirichlet distribution.