step_instance_hardness: Remove hard to classify points

In themis: Extra Recipes Steps for Dealing with Unbalanced Data

Remove hard to classify points

Description

step_instance_hardness() creates a specification of a recipe step that removes majority class instances by under-sampling the points that are hardest to classify.

Usage

step_instance_hardness(
  recipe,
  ...,
  role = NA,
  trained = FALSE,
  column = NULL,
  under_ratio = 1,
  neighbors = 5,
  distance = "euclidean",
  skip = TRUE,
  seed = sample.int(10^5, 1),
  distance_with = recipes::all_predictors(),
  id = rand_id("instance_hardness")
)

Arguments

recipe	A recipe object. The step will be added to the sequence of operations for this recipe.
...	One or more selector functions to choose which variable is used to sample the data. See recipes::selections for more details. The selection should result in single factor variable. For the tidy method, these are not currently used.
role	Not used by this step since no new variables are created.
trained	A logical to indicate if the quantities for preprocessing have been estimated.
column	A character string of the variable name that will be populated (eventually) by the ... selectors.
under_ratio	A numeric value for the ratio of the majority-to-minority frequencies. The default value (1) means that all other levels are sampled down to have the same frequency as the least occurring level. A value of 2 would mean that the majority levels will have (at most) (approximately) twice as many rows than the minority level. A named numeric vector can be used instead to give different levels different targets, for example c(a = 2, b = 3). The names must be levels of the outcome and the values are ratios of the minority level, exactly as in the single-number case. Levels that are not named are left untouched, as are rows with a missing outcome. Because a vector of targets is not a single value, supplying one means this argument can no longer be tuned. See vignette("ratio", package = "themis") for more details.
neighbors	An integer. Number of nearest neighbor that are used to generate the new examples of the minority class.
distance	A character string specifying the distance metric used for nearest neighbor calculations, defaulting to "euclidean". The available metrics fall into three groups. "euclidean", "cosine", and "mahalanobis" use approximate nearest neighbors via the RANN package and scale well to large datasets. "squared_chord", "matusita", "hellinger", and "bhattacharyya" are probability-divergence measures that treat each row as a distribution over the predictors, so they require non-negative values. "hellinger" and "bhattacharyya" further require each row to sum to 1. All four also use the RANN package and scale well to large datasets. "manhattan", "chebyshev", "canberra", "soergel", "lorentzian", "jeffreys", "topsoe", "jensen-shannon", "jensen_difference", "taneja", and "kumar-johnson" compute an exact all-pairs distance matrix. This takes time and memory proportional to the square of the number of observations in a class, so these are best suited to smaller datasets. Everything from "canberra" onwards is a probability divergence requiring non-negative values, is provided by the philentropy package (which must be installed separately), and in the case of "jeffreys", "taneja", and "kumar-johnson" requires strictly positive values, since those divide by individual predictor values. The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors.
skip	A logical. Should the step be skipped when the recipe is baked by bake()? While all operations are baked when prep() is run, some operations may not be able to be conducted on new data (e.g. processing the outcome variable(s)). Care should be taken when using skip = TRUE as it may affect the computations for subsequent operations.
seed	An integer that will be used as the seed when applied.
distance_with	A call to a selector function to choose which variables are used for distance calculations. Defaults to recipes::all_predictors(). The variable selected by ... is always excluded from the distance calculations.
id	A character string that is unique to this step to identify it.

Details

The instance hardness of each observation is estimated using the k-Disagreeing Neighbors measure: the proportion of the nearest neighbors that belong to a different class. Observations that are surrounded by points of a different class are considered hard to classify. For each majority class, the hardest observations are removed until the desired under_ratio is reached.

All columns in the data are sampled and returned by recipes::juice() and recipes::bake().

All columns selected by distance_with must be numeric with no missing data.

When used in modeling, users should strongly consider using the option skip = TRUE so that the extra sampling is not conducted outside of the training set.

Value

An updated version of recipe with the new step added to the sequence of existing steps (if any). For the tidy method, a tibble with columns terms which is the variable used to sample.

Minimum observations

The data must have at least neighbors + 1 observations for the nearest neighbors to be computed.

Tidying

When you tidy() this step, a tibble is returned with columns terms and id:

terms

character, the selectors or variables selected

id

character, id of this step

Tuning Parameters

This step has 2 tuning parameters:

• under_ratio: Under-Sampling Ratio (type: double, default: 1)

• neighbors: # Nearest Neighbors (type: integer, default: 5)

Case weights

The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.

References

Smith, M. R., Martinez, T., & Giraud-Carrier, C. (2014). An instance level analysis of data complexity. Machine learning, 95(2), 225-256.

See Also

instance_hardness() for direct implementation

Other Steps for under-sampling: step_cluster_centroids(), step_cnn(), step_downsample(), step_enn(), step_ncl(), step_nearmiss(), step_oss(), step_tomek()

Examples

library(recipes)
library(modeldata)
data(hpc_data)

hpc_data0 <- hpc_data |>
  select(-protocol, -day)

orig <- count(hpc_data0, class, name = "orig")
orig

up_rec <- recipe(class ~ ., data = hpc_data0) |>
  # Bring the majority levels down to about 1000 each
  # 1000/259 is approx 3.862
  step_instance_hardness(class, under_ratio = 3.862) |>
  prep()

training <- up_rec |>
  bake(new_data = NULL) |>
  count(class, name = "training")
training

# Since `skip` defaults to TRUE, baking the step has no effect
baked <- up_rec |>
  bake(new_data = hpc_data0) |>
  count(class, name = "baked")
baked

# Note that if the original data contained fewer rows than the
# target n (= ratio * minority_n), the data are left alone:
orig |>
  left_join(training, by = "class") |>
  left_join(baked, by = "class")

library(ggplot2)

ggplot(circle_example, aes(x, y, color = class)) +
  geom_point() +
  labs(title = "Without instance hardness") +
  xlim(c(1, 15)) +
  ylim(c(1, 15))

recipe(class ~ x + y, data = circle_example) |>
  step_instance_hardness(class) |>
  prep() |>
  bake(new_data = NULL) |>
  ggplot(aes(x, y, color = class)) +
  geom_point() +
  labs(title = "With instance hardness") +
  xlim(c(1, 15)) +
  ylim(c(1, 15))

themis documentation built on Aug. 2, 2026, 9:07 a.m.