ggRandomForests (version 3.1.2)

plot.gg_beta_varpro: Plot a gg_beta_varpro object

Description

Horizontal bar chart of the mean absolute coefficient \(\mathrm{mean}(|\hat{\beta}|)\) per variable, sorted descending so the eye lands on the top variable first. Bars filled blue when above the selection cutoff, grey otherwise. Dashed red line marks the cutoff.

Usage

# S3 method for gg_beta_varpro
plot(x, ...)

Value

A ggplot object.

Arguments

x

A gg_beta_varpro object from gg_beta_varpro().

...

Not currently used.

Reading the chart

Each bar is the average magnitude of a per-rule lasso coefficient for that variable. The numeric scale carries the predictor's units. If "age" is in years and "creatinine" is in mg/dL, a longer bar for age does not mean age is "more important" in any unit-free sense. Comparisons across data sets or across variables with very different units require keeping the units context in mind. Within one data set, bars are comparable up to that unit caveat.

Variables above the cutoff are coloured blue and flagged selected; variables below are grey. Lasso shrinkage can drive a rule's \(\hat{\beta}\) to exactly zero; those rules are kept in the average, so a variable with many shrunk-to-zero rules will sit lower in the ranking than one whose released coefficients are consistently non-zero.

For a classification fit, variables are sorted by mean(|sum-of-class-beta|) descending and that ordering is shared across every facet, so rows line up between classes for visual comparison. Each facet has its own cutoff line.

What this tells you

Use the bar chart as a selection ranking, not as an effect-size axis. Pair it with gg_varpro() to see where split-strength importance and local lasso-beta importance agree or disagree; disagreement is often the interesting signal.

See Also

gg_beta_varpro().

Examples

# \donttest{
if (requireNamespace("varPro", quietly = TRUE)) {
  set.seed(1)
  v <- varPro::varpro(mpg ~ ., data = mtcars, ntree = 50)
  plot(gg_beta_varpro(v))
}
# }