Skip to contents

Compare learners or hyperparameter grids on PU data

Source: R/xplus_grid.R
xplus_grid.Rd

The model-selection piece of the multi-backend API: fit xplus() for every candidate and rank them by a PU-valid pu_score() — no negative labels required. learners accepts a character vector of built-in backend names, a list mixing names and xplus_backend() objects, or a single constructor function (e.g. xplus_backend_xgboost) combined with params, a data frame whose columns are constructor arguments and whose rows form the tuning grid. One shared seed gives every candidate the same bootstrap draws, so the comparison is paired rather than noisy. Failures are captured per candidate instead of aborting the grid.

Usage

xplus_grid(x, y, learners, params = NULL, metric = "pu_auc", seed = NULL, ...)

Arguments

x

Feature matrix as in xplus().

y

Binary PU labels: 1 positive, 0 unlabeled.

learners

Character vector of backend names, a list of backend names/specs, or a backend constructor function to combine with params.

params

Optional data frame of constructor arguments when learners is a function; each row is one candidate.

metric

Scoring metric passed to pu_score().

seed

Optional seed shared by every candidate fit.

...

Additional arguments passed to xplus() for every candidate.

Value

A data frame with one row per candidate: learner, any params columns, score, n_iter, stop_reason, seconds, and error (NA on success). Sorted by descending score.

See also

xplus(), pu_score(), xplus_backend()

Examples

sim <- simulate_pu_data(n = 300, p = 6, boundary = "nonlinear", seed = 1,
                        prevalence = 0.4, label_propensity = 0.5)
xplus_grid(sim$x, sim$y, c("glmnet", "ranger"), max_iter = 5, seed = 1)
#>   learner     score n_iter stop_reason seconds error
#> 2  ranger 0.7580417      5    max_iter   0.322  <NA>
#> 1  glmnet 0.5981303      5    max_iter   0.078  <NA>