Skip to contents

PU learning baselines on the xplus backend contract

Source: R/pu_baselines.R
xplus_elkan_noto.Rd

Classical positive-unlabeled algorithms expressed on the same backend contract as xplus(), for honest head-to-head comparisons.

Usage

xplus_elkan_noto(
  x,
  y,
  backend = "glmnet",
  seed = NULL,
  nfolds = 10,
  estimate = c("mean", "median"),
  alpha = 1,
  cv_measure = "deviance"
)

xplus_naive(
  x,
  y,
  backend = "glmnet",
  seed = NULL,
  nfolds = 10,
  alpha = 1,
  cv_measure = "deviance"
)

Arguments

x

Finite numeric feature matrix with one row per observation.

y

Binary PU labels: 1 labeled positive, 0 unlabeled.

backend

A backend name or xplus_backend() object, resolved exactly as in xplus().

seed

Optional integer seed for reproducible resampling.

nfolds

Folds passed to the backend fit contract (used by CV backends for honest predictions).

estimate

How to estimate the labeling propensity c: "mean" (Elkan & Noto's validation-set average) or "median" over labeled positives, which is more robust to contaminated positives.

alpha, cv_measure

glmnet backend configuration, as in xplus().

Value

An xplus_baseline object with predict support, $cutoff (0.5 on the rescaled scale for Elkan-Noto), $scale (estimated c), and $method.

Details

xplus_elkan_noto() implements the Elkan & Noto (2008) recipe: train a classifier separating labeled positives (s = 1) from the unlabeled pool, estimate the labeling propensity c = P(s = 1 | y = 1) from held-out-style predictions on labeled positives, and rescale p(y = 1 | x) = p(s = 1 | x) / c clipped to one. xplus_naive() is the lower bound: a supervised classifier trained treating all unlabeled instances as negatives.

References

Elkan, C. and Noto, K. (2008). Learning classifiers from only positive and unlabeled data. KDD.

Examples

set.seed(1)
sim <- simulate_pu_data(n = 200, p = 5, prevalence = 0.4,
                        label_propensity = 0.5)
fit <- xplus_elkan_noto(sim$x, sim$y, backend = "ranger")
head(predict(fit, newx = sim$x))
#> [1] 0.27426037 0.35425298 1.00000000 1.00000000 0.08570637 0.38282177