PU-valid score for a fitted xplus object

Rank learners without negative labels. The only honest supervision in positive-unlabeled data is the labeled-positive vs unlabeled contrast, so the primary score is the AUC of the stored honest predictions (object$pred_y) with unlabeled cases treated as negatives. Under SCAR labeling this proxy is biased toward conservatism but ranks learners consistently — the standard model-selection trick of the PU literature.

Usage

pu_score(object, metric = c("pu_auc", "stability", "pseudo_auc"))

Arguments

object

A fitted "xplus" object.

metric

"pu_auc" (positive-vs-unlabeled AUC on honest scores), "stability" (mean pseudo-label stability score across iterations), or "pseudo_auc" (AUC of honest scores against the soft final labels).

Value

A numeric scalar; higher is better.

See also

xplus(), xplus_grid()

Examples

fit <- xplus(matrix(rnorm(100 * 4), ncol = 4),
             c(rep(1, 20), rep(0, 80)), max_iter = 3)
pu_score(fit)
#> [1] 0.6025