Simulate a positive-unlabeled dataset

Generate a binary dataset with known ground truth for benchmarking xplus() backends. Features are independent standard normals; the true positive probability follows a configurable decision boundary. Positive cases are observed under the SCAR assumption (selected completely at random): each true positive is labeled with constant probability label_propensity, and everything else remains unlabeled.

Usage

simulate_pu_data(
  n = 500,
  p = 10,
  boundary = c("linear", "interaction", "nonlinear", "xor", "mixture"),
  prevalence = 0.5,
  label_propensity = 0.5,
  n_informative = min(5L, p),
  sharpness = 2,
  seed = NULL
)

Arguments

n

Number of observations.

p

Number of features.

boundary

Decision boundary driving the positive probability: "linear" (sparse linear predictor), "interaction" (pure pairwise products), "nonlinear" (radial, positives outside a sphere), "xor" (checkerboard of feature-sign products, unlearnable by additive learners), or "mixture" (positives concentrated near two cluster centers, multimodal).

prevalence

Target fraction of true positives in [0, 1].

label_propensity

Constant probability that a true positive is observed as a labeled positive, in (0, 1].

n_informative

Number of features driving the boundary.

sharpness

Positive logistic scale on the centered boundary score; larger values make the boundary more deterministic.

seed

Optional nonnegative integer seed.

Value

A list of class "xplus_pu_data" with: x (feature matrix), y (observed PU labels, 1 = labeled positive, 0 = unlabeled), z (true binary labels), p_true (true positive probability per observation), and the simulation settings.

See also

xplus()

Examples

sim <- simulate_pu_data(n = 300, p = 8, boundary = "interaction", seed = 1)
fit <- xplus(sim$x, sim$y, max_iter = 5)
auc(sim$z, as.numeric(predict(fit)))
#> [1] 0.6815068