EIF estimator with per-category calibration (optimal for discrete)

Learns \(g: \{1,...,K\} \rightarrow \mathbb{R}\) such that \(g(z) = E[Y | \hat{Y} = z]\). For discrete \(\hat{Y}\), this is just the conditional mean per category

Usage

eif_discrete_point_and_ci(Y_cal, Yhat_cal, Yhat_test, alpha = 0.1)

Arguments

Y_cal

Vector of true labels on the calibration set.

Yhat_cal

Vector of surrogate predictions on the calibration set.

Yhat_test

Vector of surrogate predictions on the test set.

alpha

Miscoverage level for the confidence interval.