Skip to contents

select_features_forward() scores every candidate at every step and keeps the best; its history holds all of them. This draws the accepted variable's score at each step as the path, every other candidate's score at that step as a faint point, and the step at which the selection stopped in red. When it stopped because no candidate passed the acceptance rule (rule in select_features_forward()), the path runs one step further to the best of the rejected candidates, drawn hollow and labelled "not added", so the stop reads as a flattening rather than a cut. The picture then shows whether the last variable was a clear gain or a marginal one, and whether the runner-up would have done as well. The scores are the selection's own cross-validated criterion, optimistically biased by the selection (see "Selection score" in select_features_forward()); when a hold-out score was computed (select_on = "split") it is drawn as a separate mark at the final step and named in the caption.

Usage

# S3 method for class 'feature_selection'
plot(x, ...)

Arguments

x

The list returned by select_features_forward().

...

Not used.

Value

A ggplot object.

Details

With a learner that cannot fit an empty predictor set the first variable is added without a test and can be dropped at the end (baseline$first$dropped). The path drawn is then the one the sweep walked, with that variable hollow and labelled "dropped at the end", and the red mark is the selected set, scored without it.

See also

Examples

library(sf)
set.seed(4)
n <- 150
dat <- st_as_sf(
  data.frame(x = 5e5 + runif(n, 0, 1000), y = 5e6 + runif(n, 0, 1000),
             a = rnorm(n), b = rnorm(n), c = rnorm(n)),
  coords = c("x", "y"), crs = 32632)
dat$z <- 2 * dat$a - dat$b + rnorm(n, 0, 0.5)
fit_fn <- function(train_sf, vars)
  fit_rf_model(train_sf, "z", vars, num_trees = 80, seed = 1)
sel <- select_features_forward(dat, "z", c("a", "b", "c"), fit_fn = fit_fn,
                               k = 3, quiet = TRUE)
plot(sel)