Plot DNN weight profiles, concentration, and effective kernels

Three views of the localization weights stored on a fitted object. type = "profile" draws weight against rank with a guide at the exact finite-population support edge n - s + k for DNN fits (n - s + 1 in the paper-default k = 1 case); type = "cumulative" draws the cumulative weight share with guides at 0.5 and 0.9 and markers at the top-1/5/10 shares; type = "kernel" draws weight against stored distance — the effective data-driven kernel — with a guide at the rank-s radius. The rank-s radius is a scale diagnostic, not the DNN support radius.

Usage

plot_weights(
  object,
  type = c("profile", "cumulative", "kernel"),
  targets = NULL,
  log_weights = FALSE,
  candidates = "default",
  ...
)

Arguments

object: A fitted result object holding the weight matrix W (stored by default; type = "kernel" additionally requires stored distances via return = c("W", "dx")).
type: One of "profile", "cumulative", or "kernel".
targets: Optional integer vector of target ids to display.
log_weights: Logical (default FALSE); for rank-domain views, drop non-positive weights and use a log10 y scale (the number of dropped weights is reported in the caption).
candidates: Candidate rows to compare for grid fits. The default "default" preserves the stored weight-matrix display. Use "all" or a character vector of candidate ids with type = "profile" to compare rank-weight profiles without refitting.
...: Ignored; reserved for future options.

Value

A ggplot object.

Details

When more than eight targets are stored, panels default to a subset (first, median effective sample size, last); override with targets =. Two-scale (TDNN) fits have signed weights: a dashed zero line is added automatically whenever negative weights are shown.

Examples

if (requireNamespace("ggplot2", quietly = TRUE)) {
  set.seed(1)
  n <- 80
  x_mat <- matrix(rnorm(n * 2), n, 2)
  y_vec <- x_mat[, 1] + rnorm(n, sd = 0.3)
  fit <- np_dnn(
    X = x_mat, x0 = c(0, 0), Y = y_vec,
    k = 1L, s = 12L, se = FALSE, return = c("W", "dx")
  )
  plot_weights(fit, type = "kernel")
}