Skip to contents

Table of DALEX Permutation Feature Importance

Source: R/tbl_feature_importance_dalex.R
tbl_feature_importance_dalex.Rd

Summarises the dropout loss of each variable over the permutations of a [DALEX::feature_importance()] (or `model_parts()`) result: minimum, first quartile, median, mean, third quartile and maximum, ranked by the chosen statistic. Replaces the hand-rolled aggregation that was repeated in the feature-selection child documents.

Usage

tbl_feature_importance_dalex(
  x,
  rank_by = c("mean", "median", "min", "q1", "q3", "max"),
  top_n = NULL,
  keep_reference = FALSE,
  id = NULL
)

Arguments

x

A `feature_importance_explainer` object or a data frame with the columns `variable`, `permutation` and `dropout_loss` (and optionally `label`).

rank_by

Statistic to rank by, one of `"mean"` (default), `"median"`, `"min"`, `"q1"`, `"q3"`, `"max"`. Larger dropout loss means more important.

top_n

Keep only the `top_n` highest ranked variables (default all).

keep_reference

Keep the `_baseline_` and `_full_model_` rows (default `FALSE`).

id

Optional named list of constant columns put in front, e.g. `list(index = "A")`, to stack the tables of several models with [dplyr::bind_rows()].

Value

A tibble with `rank`, `variable`, `label` (if present), `dropout_loss` (the unpermuted, `permutation == 0` value, if present) and the six summary columns.

See also

[calc_grouped_summary()]

Examples

fi <- data.frame(
  variable = rep(c("_full_model_", "a", "b", "_baseline_"), each = 4),
  permutation = rep(0:3, 4),
  dropout_loss = c(rep(0.2, 4), 0.30, 0.31, 0.29, 0.33, 0.22, 0.21, 0.23, 0.22, rep(0.5, 4)),
  label = "model"
)
tbl_feature_importance_dalex(fi)
#> # A tibble: 2 × 10
#>    rank variable label dropout_loss   min    q1 median  mean    q3   max
#>   <int> <chr>    <chr>        <dbl> <dbl> <dbl>  <dbl> <dbl> <dbl> <dbl>
#> 1     1 a        model         0.3   0.29 0.298  0.305 0.308 0.315  0.33
#> 2     2 b        model         0.22  0.21 0.218  0.22  0.22  0.222  0.23