step_kmeans_smote() creates a specification of a recipe step that
generates new examples of the minority class, restricting them to the regions
of the predictor space where that class is dominant.
step_kmeans_smote(
recipe,
...,
role = NA,
trained = FALSE,
column = NULL,
over_ratio = 1,
neighbors = 2,
num_clusters = NULL,
cluster_balance_threshold = 1,
density_exponent = NULL,
distance = "euclidean",
indicator_column = NULL,
skip = TRUE,
seed = sample.int(10^5, 1),
id = rand_id("kmeans_smote")
)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.
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.
Not used by this step since no new variables are created.
A logical to indicate if the quantities for preprocessing have been estimated.
A character string of the variable name that will
be populated (eventually) by the ... selectors.
A numeric value for the ratio of the minority-to-majority frequencies. The default value (1) means that all other levels are sampled up to have the same frequency as the most occurring level. A value of 0.5 would mean that the minority levels will have (at most) (approximately) half as many rows as the majority level.
A named numeric vector can be used instead to give different levels
different targets, for example c(a = 1, b = 0.5). The names must be
levels of the outcome and the values are ratios of the majority 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.
An integer. Number of nearest neighbor that are used to generate the new examples of the minority class. Only observations within the same cluster are considered as neighbors.
An integer, the number of clusters to split the predictor
space into, or NULL (the default) to use max(2, floor(sqrt(n / 2)))
where n is the number of observations.
A number. A cluster is used for
over-sampling a class only if it contains at least
cluster_balance_threshold times as many observations of that class as of
all other classes combined. Defaults to 1, meaning that the class must be
at least as common as the rest of the data within the cluster. Smaller
values keep more clusters.
A number, the exponent applied to the average
pairwise distance when measuring how sparse a cluster is, or NULL (the
default) to use the number of predictors.
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.
A single string or NULL (the default). If a
string is given, a logical column with that name is added to the output,
marking rows added by the step (TRUE) vs rows from the original data
(FALSE).
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.
An integer that will be used as the seed when applied.
A character string that is unique to this step to identify it.
Each minority class must have more than neighbors observations in at least
one kept cluster. If no cluster qualifies, an error is thrown suggesting
which arguments to loosen.
When you tidy() this step, a tibble is returned with
columns terms and id:
character, the selectors or variables selected
character, id of this step
This step has 3 tuning parameters:
over_ratio: Over-Sampling Ratio (type: double, default: 1)
neighbors: # Nearest Neighbors (type: integer, default: 2)
num_clusters: # Clusters (type: integer, default: NULL)
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
KMeans-SMOTE combines k-means clustering with SMOTE to avoid generating synthetic points in regions where the minority class is not actually present. It works in three stages:
The predictor space of the whole data set is clustered with
stats::kmeans() into num_clusters clusters.
Each cluster is either kept or filtered out. A cluster is kept for a
given minority class if the ratio of that class's observations to all
other observations in the cluster is at least
cluster_balance_threshold, and if the cluster contains more than
neighbors observations of that class so that interpolation is
possible.
The synthetic points are distributed over the kept clusters according to
how sparse each cluster is. The sparsity of a cluster is
mean_pairwise_distance ^ density_exponent / n, where n is the number
of minority observations in the cluster. Sparser clusters receive more
points. Within each cluster the points are then generated by ordinary
SMOTE interpolation, using only that cluster's observations as
neighbors.
Filtering on cluster balance keeps synthetic points away from majority-dominated regions, and the density weighting counteracts within-class imbalance rather than only between-class imbalance.
The clustering is done once on all observations, so in a multi-class
problem every minority class shares the same clusters; only the filtering
and weighting are done per class. Note also that k-means always uses
Euclidean distance. The distance argument affects the nearest-neighbor
search used for interpolation and the sparsity calculation, not the
clustering itself.
All columns in the data are sampled and returned by recipes::juice()
and recipes::bake().
All columns used in this step 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.
Douzas, G., Bacao, F., and Last, F. (2018). Improving imbalanced learning through a heuristic oversampling method based on k-means and SMOTE. Information Sciences, 465:1-20.
kmeans_smote() for direct implementation
step_smote() for the same interpolation without the clustering step
Other Steps for over-sampling:
step_adasyn(),
step_bsmote(),
step_rose(),
step_smogn(),
step_smote(),
step_smoten(),
step_smotenc(),
step_svmsmote(),
step_upsample()
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 minority levels up to about 1000 each
# 1000/2211 is approx 0.4523
step_kmeans_smote(class, over_ratio = 0.4523) |>
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
library(ggplot2)
ggplot(circle_example, aes(x, y, color = class)) +
geom_point() +
labs(title = "Without KMeans-SMOTE")
recipe(class ~ x + y, data = circle_example) |>
step_kmeans_smote(class) |>
prep() |>
bake(new_data = NULL) |>
ggplot(aes(x, y, color = class)) +
geom_point() +
labs(title = "With KMeans-SMOTE")
Run the code above in your browser using DataLab