RAE#

mlquantify.metrics.RAE(prev_real, prev_pred, eps=0.0)[source]#

Compute the relative absolute error between the real and predicted prevalences.

The relative absolute error is the per-class absolute error divided by the true prevalence, averaged over classes: \(\frac{1}{n}\sum_i |\hat{p}_i - p_i| / p_i\).

Parameters:

prev_realarray-like of shape (n_classes,): True prevalence values for each class.
prev_predarray-like of shape (n_classes,): Predicted prevalence values for each class.
epsfloat, default=0.0: Additive (Laplace) smoothing applied to both prevalence vectors as (p + eps) / (1 + n_classes * eps) before the division. Evaluation protocols such as APP produce samples with absent classes (zero true prevalence); smoothing – e.g. eps = 1 / (2 * sample_size) (Sebastiani, 2020) – keeps the relative error finite. With eps=0 a class with zero true prevalence yields inf.

Returns:

errorfloat: Relative absolute error across all classes.