| Title: | Extra Recipes Steps for Dealing with Unbalanced Data |
|---|---|
| Description: | A dataset with an uneven number of cases in each class is said to be unbalanced. Many models produce a subpar performance on unbalanced datasets. A dataset can be balanced by increasing the number of minority cases using SMOTE 2011 <doi:10.48550/arXiv.1106.1813>, BorderlineSMOTE 2005 <doi:10.1007/11538059_91> and ADASYN 2008. Or by decreasing the number of majority cases using NearMiss 2003 <https://www.site.uottawa.ca/~nat/Workshop2003/jzhang.pdf> or Tomek link removal 1976. |
| Authors: | Emil Hvitfeldt [aut, cre] (ORCID: <https://orcid.org/0000-0002-0679-1945>), Posit Software, PBC [cph, fnd] (ROR: <https://ror.org/03wc8by49>) |
| Maintainer: | Emil Hvitfeldt <[email protected]> |
| License: | MIT + file LICENSE |
| Version: | 1.1.0.9000 |
| Built: | 2026-09-23 11:34:56 UTC |
| Source: | https://github.com/tidymodels/themis |
Generates synthetic positive instances using ADASYN algorithm.
adasyn(df, var, k = 5, over_ratio = 1, distance = "euclidean")adasyn(df, var, k = 5, over_ratio = 1, distance = "euclidean")
df |
data.frame or tibble. Must have 1 factor variable and remaining numeric variables. |
var |
Character, name of variable containing factor variable. |
k |
An integer. Number of nearest neighbor that are used to generate the new examples of the minority class. |
over_ratio |
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 |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
ADASYN is an extension of SMOTE that adaptively generates synthetic minority class examples based on the local distribution of each minority instance. Instead of generating the same number of synthetic examples for every minority instance, ADASYN generates more synthetic examples for instances that are harder to learn, specifically those surrounded by more majority class neighbors. This focuses synthetic data generation on the difficult boundary regions of the minority class, resulting in a more informative augmentation than standard SMOTE.
All columns used in this function must be numeric with no missing data.
A data.frame or tibble, depending on type of df.
He, H., Bai, Y., Garcia, E. and Li, S. 2008. ADASYN: Adaptive synthetic sampling approach for imbalanced learning. Proceedings of IJCNN 2008. (IEEE World Congress on Computational Intelligence). IEEE International Joint Conference. pp.1322-1328.
step_adasyn() for step function of this method
Other Direct Implementations:
bsmote(),
cluster_centroids(),
cnn(),
enn(),
instance_hardness(),
kmeans_smote(),
ncl(),
nearmiss(),
oss(),
rose(),
smogn(),
smote(),
smoten(),
smotenc(),
svmsmote(),
tomek()
circle_numeric <- circle_example[, c("x", "y", "class")] res <- adasyn(circle_numeric, var = "class") res <- adasyn(circle_numeric, var = "class", k = 10) res <- adasyn(circle_numeric, var = "class", over_ratio = 0.8) res <- adasyn(circle_numeric, var = "class", distance = "manhattan")circle_numeric <- circle_example[, c("x", "y", "class")] res <- adasyn(circle_numeric, var = "class") res <- adasyn(circle_numeric, var = "class", k = 10) res <- adasyn(circle_numeric, var = "class", over_ratio = 0.8) res <- adasyn(circle_numeric, var = "class", distance = "manhattan")
BSMOTE generates new examples of the minority class using nearest neighbors of these cases in the border region between classes.
bsmote( df, var, k = 5, over_ratio = 1, all_neighbors = FALSE, distance = "euclidean" )bsmote( df, var, k = 5, over_ratio = 1, all_neighbors = FALSE, distance = "euclidean" )
df |
data.frame or tibble. Must have 1 factor variable and remaining numeric variables. |
var |
Character, name of variable containing factor variable. |
k |
An integer. Number of nearest neighbor that are used to generate the new examples of the minority class. |
over_ratio |
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 |
all_neighbors |
Type of two borderline-SMOTE method. Defaults to FALSE. See details. |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
BSMOTE (borderline-SMOTE) works the same way as SMOTE, except that instead of generating points around every point of the minority class each point is first classified into the boxes "danger" and "not". For each point the nearest neighbors are calculated. If all the neighbors come from a different class it is labeled noise and put into the "not" box. If more than half of the neighbors come from a different class it is labeled "danger". Points are generated around points labeled "danger".
If all_neighbors = FALSE then points are generated between nearest
neighbors in its own class. If all_neighbors = TRUE then points are
generated between any nearest neighbors. See examples for visualization.
SMOTE generates new examples of the minority class using nearest neighbors
of these cases. For each existing minority class example, new examples are
created by interpolating between the example and its nearest neighbors. The
number of nearest neighbors used is controlled by the number of neighbors
argument (k in smote(), neighbors in step_smote()), and the number
of new examples generated is controlled by over_ratio.
All columns used in this function must be numeric with no missing data.
A data.frame or tibble, depending on type of df.
Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderline-smote: a new over-sampling method in imbalanced data sets learning. In International Conference on Intelligent Computing, pages 878–887. Springer, 2005.
step_bsmote() for step function of this method
Other Direct Implementations:
adasyn(),
cluster_centroids(),
cnn(),
enn(),
instance_hardness(),
kmeans_smote(),
ncl(),
nearmiss(),
oss(),
rose(),
smogn(),
smote(),
smoten(),
smotenc(),
svmsmote(),
tomek()
circle_numeric <- circle_example[, c("x", "y", "class")] res <- bsmote(circle_numeric, var = "class") res <- bsmote(circle_numeric, var = "class", k = 10) res <- bsmote(circle_numeric, var = "class", over_ratio = 0.8) res <- bsmote(circle_numeric, var = "class", all_neighbors = TRUE) res <- bsmote(circle_numeric, var = "class", distance = "manhattan")circle_numeric <- circle_example[, c("x", "y", "class")] res <- bsmote(circle_numeric, var = "class") res <- bsmote(circle_numeric, var = "class", k = 10) res <- bsmote(circle_numeric, var = "class", over_ratio = 0.8) res <- bsmote(circle_numeric, var = "class", all_neighbors = TRUE) res <- bsmote(circle_numeric, var = "class", distance = "manhattan")
A random dataset with two classes one of which is inside a circle. Used for examples to show how the different methods handles borders.
circle_examplecircle_example
A data frame with 400 rows and 4 variables:
Numeric.
Numeric.
Factor, values "Circle" and "Rest".
character, ID variable.
Synthetic data, simulated in data-raw/circle_example.R.
Under-samples the majority classes by replacing them with cluster representatives found with k-means.
cluster_centroids(df, var, under_ratio = 1, voting = "soft")cluster_centroids(df, var, under_ratio = 1, voting = "soft")
df |
data.frame or tibble. Must have 1 factor variable and remaining numeric variables. |
var |
Character, name of variable containing factor variable. |
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 |
voting |
A character string. Either |
Each class larger than the target count is summarized by running k-means on the observations of that class, using as many clusters as the target count. The class is then replaced by one representative per cluster:
voting = "soft"the cluster centroids themselves are used, so the returned observations are synthetic points that need not appear in the input.
voting = "hard"the observation closest to each centroid is used, so all returned observations are real rows. The representative is always picked from the class being under-sampled.
This makes it the one prototype generation under-sampler in this package. The other under-sampling methods perform prototype selection, keeping a subset of the original rows.
Because two clusters can share the same closest observation, voting = "hard" can return slightly fewer observations than the target count.
All columns used in this function must be numeric with no missing data.
A data.frame or tibble, depending on type of df.
step_cluster_centroids() for step function of this method
Other Direct Implementations:
adasyn(),
bsmote(),
cnn(),
enn(),
instance_hardness(),
kmeans_smote(),
ncl(),
nearmiss(),
oss(),
rose(),
smogn(),
smote(),
smoten(),
smotenc(),
svmsmote(),
tomek()
circle_numeric <- circle_example[, c("x", "y", "class")] res <- cluster_centroids(circle_numeric, var = "class") res <- cluster_centroids(circle_numeric, var = "class", voting = "hard") res <- cluster_centroids(circle_numeric, var = "class", under_ratio = 1.5)circle_numeric <- circle_example[, c("x", "y", "class")] res <- cluster_centroids(circle_numeric, var = "class") res <- cluster_centroids(circle_numeric, var = "class", voting = "hard") res <- cluster_centroids(circle_numeric, var = "class", under_ratio = 1.5)
Under-samples the majority classes by keeping only a consistent subset of observations that correctly classifies the data using a 1-nearest-neighbor rule.
cnn(df, var, distance = "euclidean")cnn(df, var, distance = "euclidean")
df |
data.frame or tibble. Must have 1 factor variable and remaining numeric variables. |
var |
Character, name of variable containing factor variable. |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
Condensed Nearest Neighbors (CNN) is an under-sampling method that reduces the majority classes to a consistent subset: a subset that classifies the original data correctly using a 1-nearest-neighbor rule. It starts with a "store" containing all minority class observations and one randomly chosen majority class observation. It then repeatedly scans the remaining majority class observations and moves any that are misclassified by a 1-nearest neighbor fit on the current store into the store. This continues until a full pass adds no new observations. The observations left outside the store are removed.
The smallest class is treated as the minority class and is always kept. CNN tends to keep observations near the decision boundary while discarding redundant interior observations. Because the seed observation and the scan order are random, results depend on the random seed.
All columns used in this function must be numeric with no missing data.
A data.frame or tibble, depending on type of df.
Hart, P. (1968). The condensed nearest neighbor rule. IEEE Transactions on Information Theory, 14(3), 515-516.
step_cnn() for step function of this method
Other Direct Implementations:
adasyn(),
bsmote(),
cluster_centroids(),
enn(),
instance_hardness(),
kmeans_smote(),
ncl(),
nearmiss(),
oss(),
rose(),
smogn(),
smote(),
smoten(),
smotenc(),
svmsmote(),
tomek()
circle_numeric <- circle_example[, c("x", "y", "class")] res <- cnn(circle_numeric, var = "class") res <- cnn(circle_numeric, var = "class", distance = "manhattan")circle_numeric <- circle_example[, c("x", "y", "class")] res <- cnn(circle_numeric, var = "class") res <- cnn(circle_numeric, var = "class", distance = "manhattan")
Removes observations whose class differs from the majority of their nearest neighbors.
enn( df, var, neighbors = 3, distance = "euclidean", times = 1, all_k = FALSE, kind_sel = "mode" )enn( df, var, neighbors = 3, distance = "euclidean", times = 1, all_k = FALSE, kind_sel = "mode" )
df |
data.frame or tibble. Must have 1 factor variable and remaining numeric variables. |
var |
Character, name of variable containing factor variable. |
neighbors |
An integer. Number of nearest neighbor that are used to
decide whether an observation is removed. Defaults to |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
times |
A positive integer for the maximum number of times ENN is
applied. Defaults to |
all_k |
A logical. When |
kind_sel |
A character string. The rule used to decide whether an
observation is removed. |
Edited Nearest Neighbors (ENN) is a cleaning method. For each observation it
finds the neighbors nearest neighbors and, if the class of the observation
does not match the majority class among those neighbors, the observation is
removed. This tends to remove noisy and borderline observations, which can
lead to smoother decision boundaries.
Setting times greater than 1 applies ENN repeatedly, removing more noisy
and borderline observations on each pass and stopping early once a pass
removes nothing. This corresponds to Repeated Edited Nearest Neighbors
(RENN).
Setting all_k = TRUE applies ENN with increasing numbers of neighbors, from
1 up to neighbors, cleaning the data at each step. This corresponds to
All k-Nearest Neighbors (AllKNN) and takes precedence over times.
Setting kind_sel = "all" uses a stricter cleaning rule: instead of removing
an observation when the majority of its neighbors disagree, it is removed
unless every one of its neighbors shares its class. This removes more
observations than the default kind_sel = "mode".
All columns used in this function must be numeric with no missing data.
Use times = Inf to repeat ENN until convergence.
A data.frame or tibble, depending on type of df.
Wilson, D. L. (1972). Asymptotic properties of nearest neighbor rules using edited data. IEEE Transactions on Systems, Man, and Cybernetics, (3), 408-421.
Tomek, I. (1976). An experiment with the edited nearest-neighbor rule. IEEE Transactions on Systems, Man, and Cybernetics, (6), 448-452.
step_enn() for step function of this method
Other Direct Implementations:
adasyn(),
bsmote(),
cluster_centroids(),
cnn(),
instance_hardness(),
kmeans_smote(),
ncl(),
nearmiss(),
oss(),
rose(),
smogn(),
smote(),
smoten(),
smotenc(),
svmsmote(),
tomek()
circle_numeric <- circle_example[, c("x", "y", "class")] res <- enn(circle_numeric, var = "class") res <- enn(circle_numeric, var = "class", neighbors = 5) res <- enn(circle_numeric, var = "class", distance = "manhattan") # Repeated Edited Nearest Neighbors (RENN) res <- enn(circle_numeric, var = "class", times = Inf) # All k-Nearest Neighbors (AllKNN) res <- enn(circle_numeric, var = "class", all_k = TRUE) # Stricter cleaning: remove unless all neighbors agree res <- enn(circle_numeric, var = "class", kind_sel = "all")circle_numeric <- circle_example[, c("x", "y", "class")] res <- enn(circle_numeric, var = "class") res <- enn(circle_numeric, var = "class", neighbors = 5) res <- enn(circle_numeric, var = "class", distance = "manhattan") # Repeated Edited Nearest Neighbors (RENN) res <- enn(circle_numeric, var = "class", times = Inf) # All k-Nearest Neighbors (AllKNN) res <- enn(circle_numeric, var = "class", all_k = TRUE) # Stricter cleaning: remove unless all neighbors agree res <- enn(circle_numeric, var = "class", kind_sel = "all")
Under-samples the majority classes by removing the points that are hardest to classify.
instance_hardness(df, var, k = 5, under_ratio = 1, distance = "euclidean")instance_hardness(df, var, k = 5, under_ratio = 1, distance = "euclidean")
df |
data.frame or tibble. Must have 1 factor variable and remaining numeric variables. |
var |
Character, name of variable containing factor variable. |
k |
An integer. Number of nearest neighbors used to estimate the instance hardness of each observation. |
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 |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
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 used in this function must be numeric with no missing data.
A data.frame or tibble, depending on type of df.
Smith, M. R., Martinez, T., & Giraud-Carrier, C. (2014). An instance level analysis of data complexity. Machine learning, 95(2), 225-256.
step_instance_hardness() for step function of this method
Other Direct Implementations:
adasyn(),
bsmote(),
cluster_centroids(),
cnn(),
enn(),
kmeans_smote(),
ncl(),
nearmiss(),
oss(),
rose(),
smogn(),
smote(),
smoten(),
smotenc(),
svmsmote(),
tomek()
circle_numeric <- circle_example[, c("x", "y", "class")] res <- instance_hardness(circle_numeric, var = "class") res <- instance_hardness(circle_numeric, var = "class", k = 10) res <- instance_hardness(circle_numeric, var = "class", under_ratio = 1.5) res <- instance_hardness(circle_numeric, var = "class", distance = "manhattan")circle_numeric <- circle_example[, c("x", "y", "class")] res <- instance_hardness(circle_numeric, var = "class") res <- instance_hardness(circle_numeric, var = "class", k = 10) res <- instance_hardness(circle_numeric, var = "class", under_ratio = 1.5) res <- instance_hardness(circle_numeric, var = "class", distance = "manhattan")
KMeans-SMOTE clusters the predictor space, keeps only the clusters that are dominated by the class being over-sampled, and generates new examples with SMOTE inside those clusters, giving sparser clusters more of the new points.
kmeans_smote( df, var, k = 2, over_ratio = 1, num_clusters = NULL, cluster_balance_threshold = 1, density_exponent = NULL, distance = "euclidean" )kmeans_smote( df, var, k = 2, over_ratio = 1, num_clusters = NULL, cluster_balance_threshold = 1, density_exponent = NULL, distance = "euclidean" )
df |
data.frame or tibble. Must have 1 factor variable and remaining numeric variables. |
var |
Character, name of variable containing factor variable. |
k |
An integer. Number of nearest neighbor that are used to generate the new examples of the minority class. |
over_ratio |
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 |
num_clusters |
An integer, the number of clusters to split the predictor
space into, or |
cluster_balance_threshold |
A number. A cluster is used for
over-sampling a class only if it contains at least
|
density_exponent |
A number, the exponent applied to the average
pairwise distance when measuring how sparse a cluster is, or |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
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 used in this function must be numeric with no missing data.
A data.frame or tibble, depending on type of df.
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.
step_kmeans_smote() for step function of this method
Other Direct Implementations:
adasyn(),
bsmote(),
cluster_centroids(),
cnn(),
enn(),
instance_hardness(),
ncl(),
nearmiss(),
oss(),
rose(),
smogn(),
smote(),
smoten(),
smotenc(),
svmsmote(),
tomek()
circle_numeric <- circle_example[, c("x", "y", "class")] res <- kmeans_smote(circle_numeric, var = "class") res <- kmeans_smote(circle_numeric, var = "class", num_clusters = 10) res <- kmeans_smote(circle_numeric, var = "class", over_ratio = 0.8)circle_numeric <- circle_example[, c("x", "y", "class")] res <- kmeans_smote(circle_numeric, var = "class") res <- kmeans_smote(circle_numeric, var = "class", num_clusters = 10) res <- kmeans_smote(circle_numeric, var = "class", over_ratio = 0.8)
Under-samples the majority classes by cleaning noisy observations and observations that pollute the neighborhood of minority class observations.
ncl(df, var, neighbors = 3, distance = "euclidean", threshold_clean = 0.5)ncl(df, var, neighbors = 3, distance = "euclidean", threshold_clean = 0.5)
df |
data.frame or tibble. Must have 1 factor variable and remaining numeric variables. |
var |
Character, name of variable containing factor variable. |
neighbors |
An integer. Number of nearest neighbor that are used to
decide whether an observation is removed. Defaults to |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
threshold_clean |
A numeric. Majority classes are only cleaned around
minority class observations when their size is greater than
|
The Neighborhood Cleaning Rule (NCL) is a cleaning method that combines two
passes over the data. First, it applies the Edited Nearest Neighbors rule,
removing majority class observations whose class differs from the majority of
their neighbors nearest neighbors. Second, for each minority class
observation that is itself misclassified by its neighbors, the majority class
observations among those neighbors are removed. Compared to Edited Nearest
Neighbors, this focuses the cleaning on the neighborhoods of minority class
observations.
The smallest class is treated as the minority class. Only majority classes
larger than threshold_clean times the size of the minority class are
cleaned in the second pass.
All columns used in this function must be numeric with no missing data.
A data.frame or tibble, depending on type of df.
Laurikkala, J. (2001). Improving identification of difficult small classes by balancing class distribution. In Conference on Artificial Intelligence in Medicine in Europe (pp. 63-66). Springer.
step_ncl() for step function of this method
Other Direct Implementations:
adasyn(),
bsmote(),
cluster_centroids(),
cnn(),
enn(),
instance_hardness(),
kmeans_smote(),
nearmiss(),
oss(),
rose(),
smogn(),
smote(),
smoten(),
smotenc(),
svmsmote(),
tomek()
circle_numeric <- circle_example[, c("x", "y", "class")] res <- ncl(circle_numeric, var = "class") res <- ncl(circle_numeric, var = "class", neighbors = 5) res <- ncl(circle_numeric, var = "class", distance = "manhattan")circle_numeric <- circle_example[, c("x", "y", "class")] res <- ncl(circle_numeric, var = "class") res <- ncl(circle_numeric, var = "class", neighbors = 5) res <- ncl(circle_numeric, var = "class", distance = "manhattan")
Generates synthetic positive instances using nearmiss algorithm.
nearmiss( df, var, k = 5, under_ratio = 1, distance = "euclidean", version = 1, n_neighbors_ver3 = 3 )nearmiss( df, var, k = 5, under_ratio = 1, distance = "euclidean", version = 1, n_neighbors_ver3 = 3 )
df |
data.frame or tibble. Must have 1 factor variable and remaining numeric variables. |
var |
Character, name of variable containing factor variable. |
k |
An integer. Number of nearest neighbor that are used to generate the new examples of the minority class. |
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 |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
version |
An integer. Which of the three NearMiss variants to use,
|
n_neighbors_ver3 |
An integer. The number of nearest neighbors used
to build the candidate pool of the NearMiss-3 variant. Only used when
|
The version argument selects between the three NearMiss variants:
version = 1Retains the points from the majority class which have the smallest mean distance to their nearest points in the minority class.
version = 2Retains the points from the majority class which have the smallest mean distance to their farthest points in the minority class.
version = 3Works in two stages. First, the n_neighbors_ver3
nearest majority class neighbors of each minority class point form a
candidate pool, and all other majority class points are removed. Then the
points of that pool which have the largest mean distance to their nearest
minority class points are retained.
Since the size of the NearMiss-3 candidate pool is governed by
n_neighbors_ver3 rather than by under_ratio, the pool can be smaller
than the target set by under_ratio. The whole pool is then retained and
the target is not reached.
With more than two classes, the mean distance is computed to the nearest points across all other classes, not only the minority class. This differs from imbalanced-learn, which measures distance to the minority class only. The binary case, the primary intended use, is unaffected.
All columns used in this function must be numeric with no missing data.
A data.frame or tibble, depending on type of df.
Inderjeet Mani and I Zhang. knn approach to unbalanced data distributions: a case study involving information extraction. In Proceedings of workshop on learning from imbalanced datasets, 2003.
step_nearmiss() for step function of this method
Other Direct Implementations:
adasyn(),
bsmote(),
cluster_centroids(),
cnn(),
enn(),
instance_hardness(),
kmeans_smote(),
ncl(),
oss(),
rose(),
smogn(),
smote(),
smoten(),
smotenc(),
svmsmote(),
tomek()
circle_numeric <- circle_example[, c("x", "y", "class")] res <- nearmiss(circle_numeric, var = "class") res <- nearmiss(circle_numeric, var = "class", k = 10) res <- nearmiss(circle_numeric, var = "class", under_ratio = 1.5) res <- nearmiss(circle_numeric, var = "class", distance = "manhattan") res <- nearmiss(circle_numeric, var = "class", version = 2) res <- nearmiss(circle_numeric, var = "class", version = 3) res <- nearmiss( circle_numeric, var = "class", version = 3, n_neighbors_ver3 = 10 )circle_numeric <- circle_example[, c("x", "y", "class")] res <- nearmiss(circle_numeric, var = "class") res <- nearmiss(circle_numeric, var = "class", k = 10) res <- nearmiss(circle_numeric, var = "class", under_ratio = 1.5) res <- nearmiss(circle_numeric, var = "class", distance = "manhattan") res <- nearmiss(circle_numeric, var = "class", version = 2) res <- nearmiss(circle_numeric, var = "class", version = 3) res <- nearmiss( circle_numeric, var = "class", version = 3, n_neighbors_ver3 = 10 )
Under-samples the majority classes by combining Condensed Nearest Neighbors and Tomek's links, first reducing redundant majority class observations and then removing majority class observations that form Tomek links with minority class observations.
oss(df, var, distance = "euclidean")oss(df, var, distance = "euclidean")
df |
data.frame or tibble. Must have 1 factor variable and remaining numeric variables. |
var |
Character, name of variable containing factor variable. |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
One-Sided Selection (OSS) is an under-sampling method that combines two cleaning techniques. It first applies Condensed Nearest Neighbors (CNN) to reduce the majority classes to a consistent subset that correctly classifies the data using a 1-nearest-neighbor rule, discarding redundant interior observations. It then applies Tomek's links to the remaining observations, removing the majority class observations that form Tomek links with minority class observations, cleaning the decision boundary.
The smallest class is treated as the minority class and is always kept. Because the CNN step relies on a random seed observation and a random scan order, results depend on the random seed.
With more than two classes, the Tomek's links step removes both members of a majority-majority link, not only links between a majority and the minority class. The binary case, the primary intended use, is unaffected.
All columns used in this function must be numeric with no missing data.
A data.frame or tibble, depending on type of df.
Kubat, M., & Matwin, S. (1997). Addressing the curse of imbalanced training sets: one-sided selection. In ICML (Vol. 97, pp. 179-186).
step_oss() for step function of this method
Other Direct Implementations:
adasyn(),
bsmote(),
cluster_centroids(),
cnn(),
enn(),
instance_hardness(),
kmeans_smote(),
ncl(),
nearmiss(),
rose(),
smogn(),
smote(),
smoten(),
smotenc(),
svmsmote(),
tomek()
circle_numeric <- circle_example[, c("x", "y", "class")] res <- oss(circle_numeric, var = "class") res <- oss(circle_numeric, var = "class", distance = "manhattan")circle_numeric <- circle_example[, c("x", "y", "class")] res <- oss(circle_numeric, var = "class") res <- oss(circle_numeric, var = "class", distance = "manhattan")
A thin wrapper around ROSE::ROSE() that generates a synthetic balanced
sample by enlarging the feature space of minority and majority class
examples.
rose( df, var, over_ratio = 1, minority_prop = 0.5, minority_smoothness = 1, majority_smoothness = 1 )rose( df, var, over_ratio = 1, minority_prop = 0.5, minority_smoothness = 1, majority_smoothness = 1 )
df |
A data.frame or tibble. Must have 1 factor variable with exactly 2 levels and remaining numeric variables. |
var |
Character, name of variable containing the 2-level factor variable. |
over_ratio |
A numeric value for the total size of the synthetic data relative to twice the size of the majority class. Unlike the other over-sampling steps this is not a per-class target, so a named vector of ratios is not accepted here. |
minority_prop |
A numeric value between 0 and 1 for the proportion of
synthetic observations from the minority class. Defaults to 0.5, which
generates an equal split of minority and majority synthetic observations.
This parameter controls the class balance within the synthetic data,
while |
minority_smoothness |
A numeric. Shrink factor to be multiplied by the smoothing parameters to estimate the conditional kernel density of the minority class. Defaults to 1. |
majority_smoothness |
A numeric. Shrink factor to be multiplied by the smoothing parameters to estimate the conditional kernel density of the majority class. Defaults to 1. |
The factor variable used to balance around must only have 2 levels.
The ROSE algorithm works by selecting an observation belonging to class k
and generating new examples in its neighborhood, which is determined by a
smoothing matrix H_k. Smaller values of minority_smoothness and
majority_smoothness shrink the entries of H_k, producing tighter
neighborhoods. This is a cautious choice when there is a concern that
excessively large neighborhoods could blur the boundaries between classes.
This function is a thin wrapper around ROSE::ROSE(). For full details on
the underlying implementation, see that function's documentation.
A data.frame or tibble, depending on type of df.
Lunardon, N., Menardi, G., and Torelli, N. (2014). ROSE: a Package for Binary Imbalanced Learning. R Journal, 6:79–89.
Menardi, G. and Torelli, N. (2014). Training and assessing classification rules with imbalanced data. Data Mining and Knowledge Discovery, 28:92–122.
step_rose() for step function of this method
Other Direct Implementations:
adasyn(),
bsmote(),
cluster_centroids(),
cnn(),
enn(),
instance_hardness(),
kmeans_smote(),
ncl(),
nearmiss(),
oss(),
smogn(),
smote(),
smoten(),
smotenc(),
svmsmote(),
tomek()
rose(circle_example[, c("x", "y", "class")], var = "class") rose(circle_example[, c("x", "y", "class")], var = "class", over_ratio = 0.8)rose(circle_example[, c("x", "y", "class")], var = "class") rose(circle_example[, c("x", "y", "class")], var = "class", over_ratio = 0.8)
SMOGN generates new examples for imbalanced regression problems. It identifies rare regions of the continuous outcome using a relevance function, over-samples those regions with a combination of SMOTE-style interpolation and the introduction of Gaussian noise, and under-samples the common regions.
smogn( df, var, threshold = 0.5, relevance = NULL, neighbors = 5, perturbation = 0.02, distance = "euclidean" )smogn( df, var, threshold = 0.5, relevance = NULL, neighbors = 5, perturbation = 0.02, distance = "euclidean" )
df |
data.frame or tibble. Must have 1 numeric outcome variable and remaining numeric variables. |
var |
Character, name of the numeric outcome variable. |
threshold |
A number between 0 and 1. Outcome values with a relevance at
or above this value are treated as rare and over-sampled. Defaults to |
relevance |
A matrix of relevance control points, or |
neighbors |
An integer. Number of nearest neighbor that are used to generate the new examples of the rare values. |
perturbation |
A number. The magnitude of the Gaussian noise added when
generating synthetic examples in unsafe regions. Defaults to |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
SMOGN is a pre-processing approach for imbalanced regression. A relevance
function assigns each outcome value a relevance score, and values with a
relevance at or above threshold are treated as rare. The data is split
into contiguous bins of rare and common outcome values. Common bins are
under-sampled and rare bins are over-sampled toward a balanced size. New
rare examples are generated either by interpolating between an example and a
nearby neighbor (when they are close enough to be considered safe) or by
perturbing the example with Gaussian noise (when they are not), where the
amount of noise is controlled by perturbation.
By default relevance is derived automatically from the boxplot extremes of
the outcome, giving the median a relevance of 0 and the extreme values a
relevance of 1. A matrix of relevance control points can instead be supplied
through relevance, with the first column giving outcome values and the
second column their relevance.
All columns used in this function must be numeric with no missing data.
A data.frame or tibble, depending on type of df.
Branco, P., Torgo, L., and Ribeiro, R. P. (2017). SMOGN: a pre-processing approach for imbalanced regression. Proceedings of Machine Learning Research, 74:36-50.
step_smogn() for step function of this method
Other Direct Implementations:
adasyn(),
bsmote(),
cluster_centroids(),
cnn(),
enn(),
instance_hardness(),
kmeans_smote(),
ncl(),
nearmiss(),
oss(),
rose(),
smote(),
smoten(),
smotenc(),
svmsmote(),
tomek()
circle_numeric <- circle_example[, c("x", "y")] res <- smogn(circle_numeric, var = "x") res <- smogn(circle_numeric, var = "x", neighbors = 10)circle_numeric <- circle_example[, c("x", "y")] res <- smogn(circle_numeric, var = "x") res <- smogn(circle_numeric, var = "x", neighbors = 10)
SMOTE generates new examples of the minority class using nearest neighbors of these cases.
smote(df, var, k = 5, over_ratio = 1, distance = "euclidean")smote(df, var, k = 5, over_ratio = 1, distance = "euclidean")
df |
data.frame or tibble. Must have 1 factor variable and remaining numeric variables. |
var |
Character, name of variable containing factor variable. |
k |
An integer. Number of nearest neighbor that are used to generate the new examples of the minority class. |
over_ratio |
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 |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
SMOTE generates new examples of the minority class using nearest neighbors
of these cases. For each existing minority class example, new examples are
created by interpolating between the example and its nearest neighbors. The
number of nearest neighbors used is controlled by the number of neighbors
argument (k in smote(), neighbors in step_smote()), and the number
of new examples generated is controlled by over_ratio.
All columns used in this function must be numeric with no missing data.
A data.frame or tibble, depending on type of df.
Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer, W. P. (2002). Smote: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16:321-357.
step_smote() for step function of this method
Other Direct Implementations:
adasyn(),
bsmote(),
cluster_centroids(),
cnn(),
enn(),
instance_hardness(),
kmeans_smote(),
ncl(),
nearmiss(),
oss(),
rose(),
smogn(),
smoten(),
smotenc(),
svmsmote(),
tomek()
circle_numeric <- circle_example[, c("x", "y", "class")] res <- smote(circle_numeric, var = "class") res <- smote(circle_numeric, var = "class", k = 10) res <- smote(circle_numeric, var = "class", over_ratio = 0.8) res <- smote(circle_numeric, var = "class", distance = "manhattan")circle_numeric <- circle_example[, c("x", "y", "class")] res <- smote(circle_numeric, var = "class") res <- smote(circle_numeric, var = "class", k = 10) res <- smote(circle_numeric, var = "class", over_ratio = 0.8) res <- smote(circle_numeric, var = "class", distance = "manhattan")
SMOTEN generates new examples of the minority class using nearest neighbors of these cases, for data sets where all predictors are categorical.
smoten(df, var, k = 5, over_ratio = 1)smoten(df, var, k = 5, over_ratio = 1)
df |
data.frame or tibble. Must have 1 factor variable used as the outcome and remaining categorical (factor or character) variables. |
var |
Character, name of variable containing factor variable. |
k |
An integer. Number of nearest neighbor that are used to generate the new examples of the minority class. |
over_ratio |
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 |
SMOTEN is a variant of SMOTE designed for data sets where all predictors are
categorical (nominal). Since standard SMOTE relies on interpolation in
continuous space, it cannot be applied to categorical features directly.
SMOTEN instead uses the Value Difference Metric (VDM) to measure the distance
between observations. For each minority class example, new synthetic examples
are generated by taking the most common value of each predictor among its
nearest neighbors. The number of new examples generated is controlled by
over_ratio.
All columns other than var must be categorical (factor or character) with
no missing data.
A data.frame or tibble, depending on type of df.
Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer, W. P. (2002). Smote: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16:321-357.
step_smoten() for step function of this method
Other Direct Implementations:
adasyn(),
bsmote(),
cluster_centroids(),
cnn(),
enn(),
instance_hardness(),
kmeans_smote(),
ncl(),
nearmiss(),
oss(),
rose(),
smogn(),
smote(),
smotenc(),
svmsmote(),
tomek()
df <- data.frame( x = factor(sample(letters[1:3], 100, replace = TRUE)), y = factor(sample(letters[1:2], 100, replace = TRUE)), class = factor(c(rep("minority", 20), rep("majority", 80))) ) res <- smoten(df, var = "class") res <- smoten(df, var = "class", k = 10) res <- smoten(df, var = "class", over_ratio = 0.8)df <- data.frame( x = factor(sample(letters[1:3], 100, replace = TRUE)), y = factor(sample(letters[1:2], 100, replace = TRUE)), class = factor(c(rep("minority", 20), rep("majority", 80))) ) res <- smoten(df, var = "class") res <- smoten(df, var = "class", k = 10) res <- smoten(df, var = "class", over_ratio = 0.8)
SMOTENC generates new examples of the minority class using nearest neighbors of these cases, and can handle categorical variables
smotenc(df, var, k = 5, over_ratio = 1)smotenc(df, var, k = 5, over_ratio = 1)
df |
data.frame or tibble. Must have 1 factor variable and remaining numeric variables. |
var |
Character, name of variable containing factor variable. |
k |
An integer. Number of nearest neighbor that are used to generate the new examples of the minority class. |
over_ratio |
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 |
SMOTENC extends SMOTE to handle data sets with a mix of numeric and
categorical predictors. For each minority class example, new synthetic
examples are generated by interpolating between the example and its nearest
neighbors using Gower's distance. Numeric features are interpolated
continuously; categorical features take the most common value among the
neighbors. The number of new examples generated is controlled by
over_ratio.
Columns can be numeric and categorical with no missing data.
A data.frame or tibble, depending on type of df.
Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer, W. P. (2002). Smote: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16:321-357.
Gower, J. C. (1971). A general coefficient of similarity and some of its properties. Biometrics 27(4):857-871. (For the distance metric used)
step_smotenc() for step function of this method
Other Direct Implementations:
adasyn(),
bsmote(),
cluster_centroids(),
cnn(),
enn(),
instance_hardness(),
kmeans_smote(),
ncl(),
nearmiss(),
oss(),
rose(),
smogn(),
smote(),
smoten(),
svmsmote(),
tomek()
circle_numeric <- circle_example[, c("x", "y", "class")] res <- smotenc(circle_numeric, var = "class") res <- smotenc(circle_numeric, var = "class", k = 10) res <- smotenc(circle_numeric, var = "class", over_ratio = 0.8)circle_numeric <- circle_example[, c("x", "y", "class")] res <- smotenc(circle_numeric, var = "class") res <- smotenc(circle_numeric, var = "class", k = 10) res <- smotenc(circle_numeric, var = "class", over_ratio = 0.8)
step_adasyn() creates a specification of a recipe step that generates
synthetic positive instances using ADASYN algorithm.
step_adasyn( recipe, ..., role = NA, trained = FALSE, column = NULL, over_ratio = 1, neighbors = 5, distance = "euclidean", indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("adasyn") )step_adasyn( recipe, ..., role = NA, trained = FALSE, column = NULL, over_ratio = 1, neighbors = 5, distance = "euclidean", indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("adasyn") )
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 |
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 |
over_ratio |
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 |
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
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
indicator_column |
A single string or |
skip |
A logical. Should the step be skipped when the recipe is baked by
|
seed |
An integer that will be used as the seed when applied. |
id |
A character string that is unique to this step to identify it. |
ADASYN is an extension of SMOTE that adaptively generates synthetic minority class examples based on the local distribution of each minority instance. Instead of generating the same number of synthetic examples for every minority instance, ADASYN generates more synthetic examples for instances that are harder to learn, specifically those surrounded by more majority class neighbors. This focuses synthetic data generation on the difficult boundary regions of the minority class, resulting in a more informative augmentation than standard SMOTE.
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.
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.
Each minority class must have at least neighbors + 1 observations to
perform the ADASYN algorithm.
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 2 tuning parameters:
over_ratio: Over-Sampling Ratio (type: double, default: 1)
neighbors: # Nearest Neighbors (type: integer, default: 5)
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
He, H., Bai, Y., Garcia, E. and Li, S. 2008. ADASYN: Adaptive synthetic sampling approach for imbalanced learning. Proceedings of IJCNN 2008. (IEEE World Congress on Computational Intelligence). IEEE International Joint Conference. pp.1322-1328.
adasyn() for direct implementation
Other Steps for over-sampling:
step_bsmote(),
step_kmeans_smote(),
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_adasyn(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 # Note that if the original data contained more rows than the # target n (= ratio * majority_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 ADASYN") recipe(class ~ x + y, data = circle_example) |> step_adasyn(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With ADASYN")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_adasyn(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 # Note that if the original data contained more rows than the # target n (= ratio * majority_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 ADASYN") recipe(class ~ x + y, data = circle_example) |> step_adasyn(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With ADASYN")
step_bsmote() creates a specification of a recipe step that generate new
examples of the minority class using nearest neighbors of these cases in the
border region between classes.
step_bsmote( recipe, ..., role = NA, trained = FALSE, column = NULL, over_ratio = 1, neighbors = 5, all_neighbors = FALSE, distance = "euclidean", indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("bsmote") )step_bsmote( recipe, ..., role = NA, trained = FALSE, column = NULL, over_ratio = 1, neighbors = 5, all_neighbors = FALSE, distance = "euclidean", indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("bsmote") )
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 |
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 |
over_ratio |
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 |
neighbors |
An integer. Number of nearest neighbor that are used to generate the new examples of the minority class. |
all_neighbors |
Type of two borderline-SMOTE method. Defaults to FALSE. See details. |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
indicator_column |
A single string or |
skip |
A logical. Should the step be skipped when the recipe is baked by
|
seed |
An integer that will be used as the seed when applied. |
id |
A character string that is unique to this step to identify it. |
BSMOTE (borderline-SMOTE) works the same way as SMOTE, except that instead of generating points around every point of the minority class each point is first classified into the boxes "danger" and "not". For each point the nearest neighbors are calculated. If all the neighbors come from a different class it is labeled noise and put into the "not" box. If more than half of the neighbors come from a different class it is labeled "danger". Points are generated around points labeled "danger".
If all_neighbors = FALSE then points are generated between nearest
neighbors in its own class. If all_neighbors = TRUE then points are
generated between any nearest neighbors. See examples for visualization.
SMOTE generates new examples of the minority class using nearest neighbors
of these cases. For each existing minority class example, new examples are
created by interpolating between the example and its nearest neighbors. The
number of nearest neighbors used is controlled by the number of neighbors
argument (k in smote(), neighbors in step_smote()), and the number
of new examples generated is controlled by over_ratio.
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.
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.
Each minority class must have at least neighbors + 1 observations to
perform the BSMOTE algorithm.
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: 5)
all_neighbors: All Neighbors (type: logical, default: FALSE)
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderline-smote: a new over-sampling method in imbalanced data sets learning. In International Conference on Intelligent Computing, pages 878–887. Springer, 2005.
bsmote() for direct implementation
Other Steps for over-sampling:
step_adasyn(),
step_kmeans_smote(),
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_bsmote(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 # Note that if the original data contained more rows than the # target n (= ratio * majority_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 SMOTE") recipe(class ~ x + y, data = circle_example) |> step_bsmote(class, all_neighbors = FALSE) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With borderline-SMOTE, all_neighbors = FALSE") recipe(class ~ x + y, data = circle_example) |> step_bsmote(class, all_neighbors = TRUE) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With borderline-SMOTE, all_neighbors = TRUE")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_bsmote(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 # Note that if the original data contained more rows than the # target n (= ratio * majority_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 SMOTE") recipe(class ~ x + y, data = circle_example) |> step_bsmote(class, all_neighbors = FALSE) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With borderline-SMOTE, all_neighbors = FALSE") recipe(class ~ x + y, data = circle_example) |> step_bsmote(class, all_neighbors = TRUE) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With borderline-SMOTE, all_neighbors = TRUE")
step_cluster_centroids() creates a specification of a recipe step that
removes majority class instances by replacing each majority class with
cluster representatives found with k-means.
step_cluster_centroids( recipe, ..., role = NA, trained = FALSE, column = NULL, under_ratio = 1, voting = "soft", skip = TRUE, seed = sample.int(10^5, 1), distance_with = recipes::all_predictors(), id = rand_id("cluster_centroids") )step_cluster_centroids( recipe, ..., role = NA, trained = FALSE, column = NULL, under_ratio = 1, voting = "soft", skip = TRUE, seed = sample.int(10^5, 1), distance_with = recipes::all_predictors(), id = rand_id("cluster_centroids") )
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 |
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 |
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 |
voting |
A character string. Either |
skip |
A logical. Should the step be skipped when the recipe is baked by
|
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 to compute the clusters. Defaults to
|
id |
A character string that is unique to this step to identify it. |
Each class larger than the target count is summarized by running k-means on the observations of that class, using as many clusters as the target count. The class is then replaced by one representative per cluster:
voting = "soft"the cluster centroids themselves are used, so the returned observations are synthetic points that need not appear in the input.
voting = "hard"the observation closest to each centroid is used, so all returned observations are real rows. The representative is always picked from the class being under-sampled.
This makes it the one prototype generation under-sampler in this package. The other under-sampling methods perform prototype selection, keeping a subset of the original rows.
Because two clusters can share the same closest observation, voting = "hard" can return slightly fewer observations than the target count.
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.
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.
With voting = "soft" the observations of an under-sampled class are
replaced by synthetic points, and columns that are not selected by
distance_with have no value for those points. Such columns are set to NA,
as they are for the over-sampling steps that synthesize observations. Use
voting = "hard" if these columns must be kept intact.
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 1 tuning parameters:
under_ratio: Under-Sampling Ratio (type: double, default: 1)
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
cluster_centroids() for direct implementation
Other Steps for under-sampling:
step_cnn(),
step_downsample(),
step_enn(),
step_instance_hardness(),
step_ncl(),
step_nearmiss(),
step_oss(),
step_tomek()
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_cluster_centroids(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 library(ggplot2) ggplot(circle_example, aes(x, y, color = class)) + geom_point() + labs(title = "Without ClusterCentroids") + xlim(c(1, 15)) + ylim(c(1, 15)) recipe(class ~ x + y, data = circle_example) |> step_cluster_centroids(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With ClusterCentroids") + xlim(c(1, 15)) + ylim(c(1, 15))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_cluster_centroids(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 library(ggplot2) ggplot(circle_example, aes(x, y, color = class)) + geom_point() + labs(title = "Without ClusterCentroids") + xlim(c(1, 15)) + ylim(c(1, 15)) recipe(class ~ x + y, data = circle_example) |> step_cluster_centroids(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With ClusterCentroids") + xlim(c(1, 15)) + ylim(c(1, 15))
step_cnn() creates a specification of a recipe step that removes
redundant majority class observations, keeping only a consistent subset that
correctly classifies the data using a 1-nearest-neighbor rule.
step_cnn( recipe, ..., role = NA, trained = FALSE, column = NULL, distance = "euclidean", skip = TRUE, seed = sample.int(10^5, 1), distance_with = recipes::all_predictors(), id = rand_id("cnn") )step_cnn( recipe, ..., role = NA, trained = FALSE, column = NULL, distance = "euclidean", skip = TRUE, seed = sample.int(10^5, 1), distance_with = recipes::all_predictors(), id = rand_id("cnn") )
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 |
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 |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
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
|
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
|
id |
A character string that is unique to this step to identify it. |
Condensed Nearest Neighbors (CNN) is an under-sampling method that reduces the majority classes to a consistent subset: a subset that classifies the original data correctly using a 1-nearest-neighbor rule. It starts with a "store" containing all minority class observations and one randomly chosen majority class observation. It then repeatedly scans the remaining majority class observations and moves any that are misclassified by a 1-nearest neighbor fit on the current store into the store. This continues until a full pass adds no new observations. The observations left outside the store are removed.
The smallest class is treated as the minority class and is always kept. CNN tends to keep observations near the decision boundary while discarding redundant interior observations. Because the seed observation and the scan order are random, results depend on the random seed.
All variables selected by distance_with must be numeric with no missing
data.
All columns in the data are sampled and returned by recipes::juice()
and recipes::bake().
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.
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.
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
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
Hart, P. (1968). The condensed nearest neighbor rule. IEEE Transactions on Information Theory, 14(3), 515-516.
cnn() for direct implementation
Other Steps for under-sampling:
step_cluster_centroids(),
step_downsample(),
step_enn(),
step_instance_hardness(),
step_ncl(),
step_nearmiss(),
step_oss(),
step_tomek()
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) |> step_cnn(class) |> 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 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 CNN") + xlim(c(1, 15)) + ylim(c(1, 15)) recipe(class ~ x + y, data = circle_example) |> step_cnn(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With CNN") + xlim(c(1, 15)) + ylim(c(1, 15))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) |> step_cnn(class) |> 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 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 CNN") + xlim(c(1, 15)) + ylim(c(1, 15)) recipe(class ~ x + y, data = circle_example) |> step_cnn(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With CNN") + xlim(c(1, 15)) + ylim(c(1, 15))
step_downsample() creates a specification of a recipe step that will
remove rows of a data set to make the occurrence of levels in a specific
factor level equal.
step_downsample( recipe, ..., under_ratio = 1, ratio = deprecated(), replacement = FALSE, role = NA, trained = FALSE, column = NULL, target = NA, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("downsample") )step_downsample( recipe, ..., under_ratio = 1, ratio = deprecated(), replacement = FALSE, role = NA, trained = FALSE, column = NULL, target = NA, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("downsample") )
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 |
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 |
ratio |
Deprecated argument; same as |
replacement |
A logical value indicating whether the
under-sample should be drawn with replacement. Defaults to
|
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 |
target |
A named numeric vector giving the number of rows to sample
each level down to. This should not be set by the user and will be
populated by |
skip |
A logical. Should the step be skipped when the recipe is baked by
|
seed |
An integer that will be used as the seed when applied. |
id |
A character string that is unique to this step to identify it. |
Down-sampling is intended to be performed on the training set
alone. For this reason, the default is skip = TRUE.
If there are missing values in the factor variable that is used to define the sampling, missing data are selected at random in the same way that the other factor levels are sampled. Missing values are not used to determine the amount of data in the minority level
For any data with factor levels occurring with the same frequency as the minority level, all data will be retained.
All columns in the data are sampled and returned by recipes::juice()
and recipes::bake().
Keep in mind that the location of down-sampling in the step may have effects. For example, if centering and scaling, it is not clear whether those operations should be conducted before or after rows are removed.
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.
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 1 tuning parameters:
under_ratio: Under-Sampling Ratio (type: double, default: 1)
This step performs an unsupervised operation that can utilize case weights.
To use them, see the documentation in recipes::case_weights and the examples on
tidymodels.org.
Other Steps for under-sampling:
step_cluster_centroids(),
step_cnn(),
step_enn(),
step_instance_hardness(),
step_ncl(),
step_nearmiss(),
step_oss(),
step_tomek()
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_downsample(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 more rows than the # target n (= ratio * majority_n), the data are left alone: orig |> left_join(training, by = "class") |> left_join(baked, by = "class") # A named vector gives each level its own target. Here only "VF" is # sampled down, to about twice the size of the minority level. recipe(class ~ ., data = hpc_data0) |> step_downsample(class, under_ratio = c(VF = 2)) |> prep() |> bake(new_data = NULL) |> count(class) library(ggplot2) ggplot(circle_example, aes(x, y, color = class)) + geom_point() + labs(title = "Without downsample") recipe(class ~ x + y, data = circle_example) |> step_downsample(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With downsample")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_downsample(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 more rows than the # target n (= ratio * majority_n), the data are left alone: orig |> left_join(training, by = "class") |> left_join(baked, by = "class") # A named vector gives each level its own target. Here only "VF" is # sampled down, to about twice the size of the minority level. recipe(class ~ ., data = hpc_data0) |> step_downsample(class, under_ratio = c(VF = 2)) |> prep() |> bake(new_data = NULL) |> count(class) library(ggplot2) ggplot(circle_example, aes(x, y, color = class)) + geom_point() + labs(title = "Without downsample") recipe(class ~ x + y, data = circle_example) |> step_downsample(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With downsample")
step_enn() creates a specification of a recipe step that removes
observations whose class differs from the majority of their nearest
neighbors.
step_enn( recipe, ..., role = NA, trained = FALSE, column = NULL, neighbors = 3, distance = "euclidean", times = 1, all_k = FALSE, kind_sel = "mode", skip = TRUE, seed = sample.int(10^5, 1), distance_with = recipes::all_predictors(), id = rand_id("enn") )step_enn( recipe, ..., role = NA, trained = FALSE, column = NULL, neighbors = 3, distance = "euclidean", times = 1, all_k = FALSE, kind_sel = "mode", skip = TRUE, seed = sample.int(10^5, 1), distance_with = recipes::all_predictors(), id = rand_id("enn") )
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 |
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 |
neighbors |
An integer. Number of nearest neighbor that are used to
decide whether an observation is removed. Defaults to |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
times |
A positive integer for the maximum number of times ENN is
applied. Defaults to |
all_k |
A logical. When |
kind_sel |
A character string. The rule used to decide whether an
observation is removed. |
skip |
A logical. Should the step be skipped when the recipe is baked by
|
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
|
id |
A character string that is unique to this step to identify it. |
Edited Nearest Neighbors (ENN) is a cleaning method. For each observation it
finds the neighbors nearest neighbors and, if the class of the observation
does not match the majority class among those neighbors, the observation is
removed. This tends to remove noisy and borderline observations, which can
lead to smoother decision boundaries.
Setting times greater than 1 applies ENN repeatedly, removing more noisy
and borderline observations on each pass and stopping early once a pass
removes nothing. This corresponds to Repeated Edited Nearest Neighbors
(RENN).
Setting all_k = TRUE applies ENN with increasing numbers of neighbors, from
1 up to neighbors, cleaning the data at each step. This corresponds to
All k-Nearest Neighbors (AllKNN) and takes precedence over times.
Setting kind_sel = "all" uses a stricter cleaning rule: instead of removing
an observation when the majority of its neighbors disagree, it is removed
unless every one of its neighbors shares its class. This removes more
observations than the default kind_sel = "mode".
All variables selected by distance_with must be numeric with no missing
data.
All columns in the data are sampled and returned by recipes::juice()
and recipes::bake().
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.
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.
The data must have at least neighbors + 1 observations for the nearest
neighbors to be computed.
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 2 tuning parameters:
neighbors: # Nearest Neighbors (type: integer, default: 3)
all_k: Pruning (type: logical, default: FALSE)
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
Wilson, D. L. (1972). Asymptotic properties of nearest neighbor rules using edited data. IEEE Transactions on Systems, Man, and Cybernetics, (3), 408-421.
Tomek, I. (1976). An experiment with the edited nearest-neighbor rule. IEEE Transactions on Systems, Man, and Cybernetics, (6), 448-452.
enn() for direct implementation
step_smote(), which is commonly composed before step_enn() to clean the
ambiguous points that over-sampling creates near the class boundary (the
equivalent of imbalanced-learn's SMOTEENN).
Other Steps for under-sampling:
step_cluster_centroids(),
step_cnn(),
step_downsample(),
step_instance_hardness(),
step_ncl(),
step_nearmiss(),
step_oss(),
step_tomek()
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) |> step_enn(class) |> 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 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 ENN") + xlim(c(1, 15)) + ylim(c(1, 15)) recipe(class ~ x + y, data = circle_example) |> step_enn(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With ENN") + xlim(c(1, 15)) + ylim(c(1, 15))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) |> step_enn(class) |> 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 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 ENN") + xlim(c(1, 15)) + ylim(c(1, 15)) recipe(class ~ x + y, data = circle_example) |> step_enn(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With ENN") + xlim(c(1, 15)) + ylim(c(1, 15))
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.
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") )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") )
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 |
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 |
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 |
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
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
|
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
|
id |
A character string that is unique to this step to identify it. |
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.
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.
The data must have at least neighbors + 1 observations for the nearest
neighbors to be computed.
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 2 tuning parameters:
under_ratio: Under-Sampling Ratio (type: double, default: 1)
neighbors: # Nearest Neighbors (type: integer, default: 5)
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
Smith, M. R., Martinez, T., & Giraud-Carrier, C. (2014). An instance level analysis of data complexity. Machine learning, 95(2), 225-256.
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()
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))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))
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") )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") )
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 |
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 |
over_ratio |
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 |
neighbors |
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. |
num_clusters |
An integer, the number of clusters to split the predictor
space into, or |
cluster_balance_threshold |
A number. A cluster is used for
over-sampling a class only if it contains at least
|
density_exponent |
A number, the exponent applied to the average
pairwise distance when measuring how sparse a cluster is, or |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
indicator_column |
A single string or |
skip |
A logical. Should the step be skipped when the recipe is baked by
|
seed |
An integer that will be used as the seed when applied. |
id |
A character string that is unique to this step to identify it. |
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.
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.
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.
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")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")
step_ncl() creates a specification of a recipe step that removes majority
class observations that are noisy or that pollute the neighborhood of
minority class observations.
step_ncl( recipe, ..., role = NA, trained = FALSE, column = NULL, neighbors = 3, distance = "euclidean", threshold_clean = 0.5, skip = TRUE, seed = sample.int(10^5, 1), distance_with = recipes::all_predictors(), id = rand_id("ncl") )step_ncl( recipe, ..., role = NA, trained = FALSE, column = NULL, neighbors = 3, distance = "euclidean", threshold_clean = 0.5, skip = TRUE, seed = sample.int(10^5, 1), distance_with = recipes::all_predictors(), id = rand_id("ncl") )
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 |
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 |
neighbors |
An integer. Number of nearest neighbor that are used to
decide whether an observation is removed. Defaults to |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
threshold_clean |
A numeric. Majority classes are only cleaned around
minority class observations when their size is greater than
|
skip |
A logical. Should the step be skipped when the recipe is baked by
|
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
|
id |
A character string that is unique to this step to identify it. |
The Neighborhood Cleaning Rule (NCL) is a cleaning method that combines two
passes over the data. First, it applies the Edited Nearest Neighbors rule,
removing majority class observations whose class differs from the majority of
their neighbors nearest neighbors. Second, for each minority class
observation that is itself misclassified by its neighbors, the majority class
observations among those neighbors are removed. Compared to Edited Nearest
Neighbors, this focuses the cleaning on the neighborhoods of minority class
observations.
The smallest class is treated as the minority class. Only majority classes
larger than threshold_clean times the size of the minority class are
cleaned in the second pass.
All variables selected by distance_with must be numeric with no missing
data.
All columns in the data are sampled and returned by recipes::juice()
and recipes::bake().
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.
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.
The data must have at least neighbors + 1 observations for the nearest
neighbors to be computed.
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 2 tuning parameters:
neighbors: # Nearest Neighbors (type: integer, default: 3)
threshold_clean: Threshold (type: double, default: 0.5)
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
Laurikkala, J. (2001). Improving identification of difficult small classes by balancing class distribution. In Conference on Artificial Intelligence in Medicine in Europe (pp. 63-66). Springer.
ncl() for direct implementation
Other Steps for under-sampling:
step_cluster_centroids(),
step_cnn(),
step_downsample(),
step_enn(),
step_instance_hardness(),
step_nearmiss(),
step_oss(),
step_tomek()
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) |> step_ncl(class) |> 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 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 NCL") + xlim(c(1, 15)) + ylim(c(1, 15)) recipe(class ~ x + y, data = circle_example) |> step_ncl(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With NCL") + xlim(c(1, 15)) + ylim(c(1, 15))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) |> step_ncl(class) |> 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 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 NCL") + xlim(c(1, 15)) + ylim(c(1, 15)) recipe(class ~ x + y, data = circle_example) |> step_ncl(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With NCL") + xlim(c(1, 15)) + ylim(c(1, 15))
step_nearmiss() creates a specification of a recipe step that removes
majority class instances by undersampling points in the majority class based
on their distance to points in the minority class.
step_nearmiss( recipe, ..., role = NA, trained = FALSE, column = NULL, under_ratio = 1, neighbors = 5, distance = "euclidean", version = 1, n_neighbors_ver3 = 3, skip = TRUE, seed = sample.int(10^5, 1), distance_with = recipes::all_predictors(), id = rand_id("nearmiss") )step_nearmiss( recipe, ..., role = NA, trained = FALSE, column = NULL, under_ratio = 1, neighbors = 5, distance = "euclidean", version = 1, n_neighbors_ver3 = 3, skip = TRUE, seed = sample.int(10^5, 1), distance_with = recipes::all_predictors(), id = rand_id("nearmiss") )
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 |
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 |
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 |
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
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
version |
An integer. Which of the three NearMiss variants to use,
|
n_neighbors_ver3 |
An integer. The number of nearest neighbors used
to build the candidate pool of the NearMiss-3 variant. Only used when
|
skip |
A logical. Should the step be skipped when the recipe is baked by
|
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
|
id |
A character string that is unique to this step to identify it. |
The version argument selects between the three NearMiss variants:
version = 1Retains the points from the majority class which have the smallest mean distance to their nearest points in the minority class.
version = 2Retains the points from the majority class which have the smallest mean distance to their farthest points in the minority class.
version = 3Works in two stages. First, the n_neighbors_ver3
nearest majority class neighbors of each minority class point form a
candidate pool, and all other majority class points are removed. Then the
points of that pool which have the largest mean distance to their nearest
minority class points are retained.
Since the size of the NearMiss-3 candidate pool is governed by
n_neighbors_ver3 rather than by under_ratio, the pool can be smaller
than the target set by under_ratio. The whole pool is then retained and
the target is not reached.
With more than two classes, the mean distance is computed to the nearest points across all other classes, not only the minority class. This differs from imbalanced-learn, which measures distance to the minority class only. The binary case, the primary intended use, is unaffected.
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.
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.
Each minority class must have at least neighbors + 1 observations to
perform the NearMiss algorithm.
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 2 tuning parameters:
under_ratio: Under-Sampling Ratio (type: double, default: 1)
neighbors: # Nearest Neighbors (type: integer, default: 5)
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
Inderjeet Mani and I Zhang. knn approach to unbalanced data distributions: a case study involving information extraction. In Proceedings of workshop on learning from imbalanced datasets, 2003.
nearmiss() for direct implementation
Other Steps for under-sampling:
step_cluster_centroids(),
step_cnn(),
step_downsample(),
step_enn(),
step_instance_hardness(),
step_ncl(),
step_oss(),
step_tomek()
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_nearmiss(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 NEARMISS") + xlim(c(1, 15)) + ylim(c(1, 15)) recipe(class ~ x + y, data = circle_example) |> step_nearmiss(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With NEARMISS") + xlim(c(1, 15)) + ylim(c(1, 15))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_nearmiss(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 NEARMISS") + xlim(c(1, 15)) + ylim(c(1, 15)) recipe(class ~ x + y, data = circle_example) |> step_nearmiss(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With NEARMISS") + xlim(c(1, 15)) + ylim(c(1, 15))
step_oss() creates a specification of a recipe step that removes
majority class observations using One-Sided Selection, combining Condensed
Nearest Neighbors and Tomek's links.
step_oss( recipe, ..., role = NA, trained = FALSE, column = NULL, distance = "euclidean", skip = TRUE, seed = sample.int(10^5, 1), distance_with = recipes::all_predictors(), id = rand_id("oss") )step_oss( recipe, ..., role = NA, trained = FALSE, column = NULL, distance = "euclidean", skip = TRUE, seed = sample.int(10^5, 1), distance_with = recipes::all_predictors(), id = rand_id("oss") )
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 |
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 |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
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
|
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
|
id |
A character string that is unique to this step to identify it. |
One-Sided Selection (OSS) is an under-sampling method that combines two cleaning techniques. It first applies Condensed Nearest Neighbors (CNN) to reduce the majority classes to a consistent subset that correctly classifies the data using a 1-nearest-neighbor rule, discarding redundant interior observations. It then applies Tomek's links to the remaining observations, removing the majority class observations that form Tomek links with minority class observations, cleaning the decision boundary.
The smallest class is treated as the minority class and is always kept. Because the CNN step relies on a random seed observation and a random scan order, results depend on the random seed.
With more than two classes, the Tomek's links step removes both members of a majority-majority link, not only links between a majority and the minority class. The binary case, the primary intended use, is unaffected.
All variables selected by distance_with must be numeric with no missing
data.
All columns in the data are sampled and returned by recipes::juice()
and recipes::bake().
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.
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.
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
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
Kubat, M., & Matwin, S. (1997). Addressing the curse of imbalanced training sets: one-sided selection. In ICML (Vol. 97, pp. 179-186).
oss() for direct implementation
Other Steps for under-sampling:
step_cluster_centroids(),
step_cnn(),
step_downsample(),
step_enn(),
step_instance_hardness(),
step_ncl(),
step_nearmiss(),
step_tomek()
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) |> step_oss(class) |> 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 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 OSS") + xlim(c(1, 15)) + ylim(c(1, 15)) recipe(class ~ x + y, data = circle_example) |> step_oss(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With OSS") + xlim(c(1, 15)) + ylim(c(1, 15))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) |> step_oss(class) |> 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 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 OSS") + xlim(c(1, 15)) + ylim(c(1, 15)) recipe(class ~ x + y, data = circle_example) |> step_oss(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With OSS") + xlim(c(1, 15)) + ylim(c(1, 15))
step_rose() creates a specification of a recipe step that generates
samples of synthetic data by enlarging the feature space of minority and
majority class examples. Using ROSE::ROSE().
step_rose( recipe, ..., role = NA, trained = FALSE, column = NULL, over_ratio = 1, minority_prop = 0.5, minority_smoothness = 1, majority_smoothness = 1, indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("rose") )step_rose( recipe, ..., role = NA, trained = FALSE, column = NULL, over_ratio = 1, minority_prop = 0.5, minority_smoothness = 1, majority_smoothness = 1, indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("rose") )
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 |
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 |
over_ratio |
A numeric value for the total size of the synthetic data relative to twice the size of the majority class. Unlike the other over-sampling steps this is not a per-class target, so a named vector of ratios is not accepted here. |
minority_prop |
A numeric value between 0 and 1 for the proportion of
synthetic observations from the minority class. Defaults to 0.5, which
generates an equal split of minority and majority synthetic observations.
This parameter controls the class balance within the synthetic data,
while |
minority_smoothness |
A numeric. Shrink factor to be multiplied by the smoothing parameters to estimate the conditional kernel density of the minority class. Defaults to 1. |
majority_smoothness |
A numeric. Shrink factor to be multiplied by the smoothing parameters to estimate the conditional kernel density of the majority class. Defaults to 1. |
indicator_column |
A single string or |
skip |
A logical. Should the step be skipped when the recipe is baked by
|
seed |
An integer that will be used as the seed when applied. |
id |
A character string that is unique to this step to identify it. |
The factor variable used to balance around must only have 2 levels.
The ROSE algorithm works by selecting an observation belonging to class k
and generating new examples in its neighborhood, which is determined by a
smoothing matrix H_k. Smaller values of minority_smoothness and
majority_smoothness shrink the entries of H_k, producing tighter
neighborhoods. This is a cautious choice when there is a concern that
excessively large neighborhoods could blur the boundaries between classes.
All columns in the data are sampled and returned by recipes::juice()
and recipes::bake().
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.
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.
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 1 tuning parameters:
over_ratio: Over-Sampling Ratio (type: double, default: 1)
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
Lunardon, N., Menardi, G., and Torelli, N. (2014). ROSE: a Package for Binary Imbalanced Learning. R Journal, 6:79–89.
Menardi, G. and Torelli, N. (2014). Training and assessing classification rules with imbalanced data. Data Mining and Knowledge Discovery, 28:92–122.
rose() for direct implementation
Other Steps for over-sampling:
step_adasyn(),
step_bsmote(),
step_kmeans_smote(),
step_smogn(),
step_smote(),
step_smoten(),
step_smotenc(),
step_svmsmote(),
step_upsample()
library(recipes) library(modeldata) data(hpc_data) hpc_data0 <- hpc_data |> mutate(class = factor(class == "VF", labels = c("not VF", "VF"))) |> select(-protocol, -day) orig <- count(hpc_data0, class, name = "orig") orig up_rec <- recipe(class ~ ., data = hpc_data0) |> step_rose(class) |> 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 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 ROSE") recipe(class ~ x + y, data = circle_example) |> step_rose(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With ROSE")library(recipes) library(modeldata) data(hpc_data) hpc_data0 <- hpc_data |> mutate(class = factor(class == "VF", labels = c("not VF", "VF"))) |> select(-protocol, -day) orig <- count(hpc_data0, class, name = "orig") orig up_rec <- recipe(class ~ ., data = hpc_data0) |> step_rose(class) |> 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 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 ROSE") recipe(class ~ x + y, data = circle_example) |> step_rose(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With ROSE")
step_smogn() creates a specification of a recipe step that generates new
examples for imbalanced regression problems using SMOGN.
step_smogn( recipe, ..., role = NA, trained = FALSE, column = NULL, threshold = 0.5, relevance = NULL, neighbors = 5, perturbation = 0.02, distance = "euclidean", indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("smogn") )step_smogn( recipe, ..., role = NA, trained = FALSE, column = NULL, threshold = 0.5, relevance = NULL, neighbors = 5, perturbation = 0.02, distance = "euclidean", indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("smogn") )
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
numeric variable. For the |
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 |
threshold |
A number between 0 and 1. Outcome values with a relevance at
or above this value are treated as rare and over-sampled. Defaults to |
relevance |
A matrix of relevance control points, or |
neighbors |
An integer. Number of nearest neighbor that are used to generate the new examples of the rare values. |
perturbation |
A number. The magnitude of the Gaussian noise added when
generating synthetic examples in unsafe regions. Defaults to |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
indicator_column |
A single string or |
skip |
A logical. Should the step be skipped when the recipe is baked by
|
seed |
An integer that will be used as the seed when applied. |
id |
A character string that is unique to this step to identify it. |
SMOGN is a pre-processing approach for imbalanced regression. A relevance
function assigns each outcome value a relevance score, and values with a
relevance at or above threshold are treated as rare. The data is split
into contiguous bins of rare and common outcome values. Common bins are
under-sampled and rare bins are over-sampled toward a balanced size. New
rare examples are generated either by interpolating between an example and a
nearby neighbor (when they are close enough to be considered safe) or by
perturbing the example with Gaussian noise (when they are not), where the
amount of noise is controlled by perturbation.
By default relevance is derived automatically from the boxplot extremes of
the outcome, giving the median a relevance of 0 and the extreme values a
relevance of 1. A matrix of relevance control points can instead be supplied
through relevance, with the first column giving outcome values and the
second column their relevance.
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.
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.
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 2 tuning parameters:
neighbors: # Nearest Neighbors (type: integer, default: 5)
threshold: Threshold (type: double, default: 0.5)
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
Branco, P., Torgo, L., and Ribeiro, R. P. (2017). SMOGN: a pre-processing approach for imbalanced regression. Proceedings of Machine Learning Research, 74:36-50.
smogn() for direct implementation
Other Steps for over-sampling:
step_adasyn(),
step_bsmote(),
step_kmeans_smote(),
step_rose(),
step_smote(),
step_smoten(),
step_smotenc(),
step_svmsmote(),
step_upsample()
library(recipes) library(ggplot2) ggplot(circle_example, aes(x)) + geom_histogram(bins = 30) + labs(title = "Without SMOGN") recipe(y ~ x, data = circle_example) |> step_smogn(y) |> prep() |> bake(new_data = NULL) |> ggplot(aes(y)) + geom_histogram(bins = 30) + labs(title = "With SMOGN")library(recipes) library(ggplot2) ggplot(circle_example, aes(x)) + geom_histogram(bins = 30) + labs(title = "Without SMOGN") recipe(y ~ x, data = circle_example) |> step_smogn(y) |> prep() |> bake(new_data = NULL) |> ggplot(aes(y)) + geom_histogram(bins = 30) + labs(title = "With SMOGN")
step_smote() creates a specification of a recipe step that generate new
examples of the minority class using nearest neighbors of these cases.
step_smote( recipe, ..., role = NA, trained = FALSE, column = NULL, over_ratio = 1, neighbors = 5, distance = "euclidean", indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("smote") )step_smote( recipe, ..., role = NA, trained = FALSE, column = NULL, over_ratio = 1, neighbors = 5, distance = "euclidean", indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("smote") )
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 |
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 |
over_ratio |
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 |
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
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
indicator_column |
A single string or |
skip |
A logical. Should the step be skipped when the recipe is baked by
|
seed |
An integer that will be used as the seed when applied. |
id |
A character string that is unique to this step to identify it. |
SMOTE generates new examples of the minority class using nearest neighbors
of these cases. For each existing minority class example, new examples are
created by interpolating between the example and its nearest neighbors. The
number of nearest neighbors used is controlled by the number of neighbors
argument (k in smote(), neighbors in step_smote()), and the number
of new examples generated is controlled by over_ratio.
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.
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.
Each minority class must have at least neighbors + 1 observations to
perform the SMOTE algorithm.
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 2 tuning parameters:
over_ratio: Over-Sampling Ratio (type: double, default: 1)
neighbors: # Nearest Neighbors (type: integer, default: 5)
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer, W. P. (2002). Smote: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16:321-357.
smote() for direct implementation
step_enn() and step_tomek(), which are commonly composed after
step_smote() to clean the ambiguous points that over-sampling creates
near the class boundary (the equivalent of imbalanced-learn's SMOTEENN
and SMOTETomek).
Other Steps for over-sampling:
step_adasyn(),
step_bsmote(),
step_kmeans_smote(),
step_rose(),
step_smogn(),
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_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 # Note that if the original data contained more rows than the # target n (= ratio * majority_n), the data are left alone: orig |> left_join(training, by = "class") |> left_join(baked, by = "class") # A named vector gives each level its own target. Here "VF" is left # untouched and only "L" is brought up to the size of the majority level. recipe(class ~ ., data = hpc_data0) |> step_smote(class, over_ratio = c(L = 1)) |> prep() |> bake(new_data = NULL) |> count(class) library(ggplot2) ggplot(circle_example, aes(x, y, color = class)) + geom_point() + labs(title = "Without SMOTE") recipe(class ~ x + y, data = circle_example) |> step_smote(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With SMOTE")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_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 # Note that if the original data contained more rows than the # target n (= ratio * majority_n), the data are left alone: orig |> left_join(training, by = "class") |> left_join(baked, by = "class") # A named vector gives each level its own target. Here "VF" is left # untouched and only "L" is brought up to the size of the majority level. recipe(class ~ ., data = hpc_data0) |> step_smote(class, over_ratio = c(L = 1)) |> prep() |> bake(new_data = NULL) |> count(class) library(ggplot2) ggplot(circle_example, aes(x, y, color = class)) + geom_point() + labs(title = "Without SMOTE") recipe(class ~ x + y, data = circle_example) |> step_smote(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With SMOTE")
step_smoten() creates a specification of a recipe step that generate new
examples of the minority class using nearest neighbors of these cases, for
data sets where all predictors are categorical (nominal). The Value
Difference Metric (VDM) is used to measure the distance between observations.
For each predictor, the most common category among neighbors is chosen.
step_smoten( recipe, ..., role = NA, trained = FALSE, column = NULL, over_ratio = 1, neighbors = 5, indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("smoten") )step_smoten( recipe, ..., role = NA, trained = FALSE, column = NULL, over_ratio = 1, neighbors = 5, indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("smoten") )
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 |
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 |
over_ratio |
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 |
neighbors |
An integer. Number of nearest neighbor that are used to generate the new examples of the minority class. |
indicator_column |
A single string or |
skip |
A logical. Should the step be skipped when the recipe is baked by
|
seed |
An integer that will be used as the seed when applied. |
id |
A character string that is unique to this step to identify it. |
SMOTEN is a variant of SMOTE designed for data sets where all predictors are
categorical (nominal). Since standard SMOTE relies on interpolation in
continuous space, it cannot be applied to categorical features directly.
SMOTEN instead uses the Value Difference Metric (VDM) to measure the distance
between observations. For each minority class example, new synthetic examples
are generated by taking the most common value of each predictor among its
nearest neighbors. The number of new examples generated is controlled by
over_ratio.
All columns in the data are sampled and returned by recipes::juice()
and recipes::bake().
All predictor columns must be categorical (factor or character) 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.
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.
Each minority class must have at least neighbors + 1 observations to
perform the SMOTEN algorithm.
The Value Difference Metric (VDM) used here deviates from Chawla's stated form in two ways. The per-feature deltas are aggregated by summing them (r = 1) rather than by taking their Euclidean norm (r = 2), and when a synthetic value is chosen by majority vote of the nearest neighbors the seed observation itself is excluded from the vote. The metric is internally consistent and is a valid VDM variant, but be aware of these choices when comparing results with other implementations.
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 2 tuning parameters:
over_ratio: Over-Sampling Ratio (type: double, default: 1)
neighbors: # Nearest Neighbors (type: integer, default: 5)
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer, W. P. (2002). Smote: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16:321-357.
smoten() for direct implementation
Other Steps for over-sampling:
step_adasyn(),
step_bsmote(),
step_kmeans_smote(),
step_rose(),
step_smogn(),
step_smote(),
step_smotenc(),
step_svmsmote(),
step_upsample()
library(recipes) library(modeldata) data(hpc_data) hpc_cat <- hpc_data[, c("class", "protocol", "day")] orig <- count(hpc_cat, class, name = "orig") orig up_rec <- recipe(class ~ ., data = hpc_cat) |> step_smoten(class) |> 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_cat) |> count(class, name = "baked") bakedlibrary(recipes) library(modeldata) data(hpc_data) hpc_cat <- hpc_data[, c("class", "protocol", "day")] orig <- count(hpc_cat, class, name = "orig") orig up_rec <- recipe(class ~ ., data = hpc_cat) |> step_smoten(class) |> 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_cat) |> count(class, name = "baked") baked
step_smotenc() creates a specification of a recipe step that generate new
examples of the minority class using nearest neighbors of these cases.
Gower's distance is used to handle mixed data types. For categorical
variables, the most common category along neighbors is chosen.
step_smotenc( recipe, ..., role = NA, trained = FALSE, column = NULL, over_ratio = 1, neighbors = 5, indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("smotenc") )step_smotenc( recipe, ..., role = NA, trained = FALSE, column = NULL, over_ratio = 1, neighbors = 5, indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("smotenc") )
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 |
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 |
over_ratio |
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 |
neighbors |
An integer. Number of nearest neighbor that are used to generate the new examples of the minority class. |
indicator_column |
A single string or |
skip |
A logical. Should the step be skipped when the recipe is baked by
|
seed |
An integer that will be used as the seed when applied. |
id |
A character string that is unique to this step to identify it. |
SMOTENC extends SMOTE to handle data sets with a mix of numeric and
categorical predictors. For each minority class example, new synthetic
examples are generated by interpolating between the example and its nearest
neighbors using Gower's distance. Numeric features are interpolated
continuously; categorical features take the most common value among the
neighbors. The number of new examples generated is controlled by
over_ratio.
All columns in the data are sampled and returned by recipes::juice()
and recipes::bake().
Columns can be numeric and categorical 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.
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.
Each minority class must have at least neighbors + 1 observations to
perform the SMOTENC algorithm.
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 2 tuning parameters:
over_ratio: Over-Sampling Ratio (type: double, default: 1)
neighbors: # Nearest Neighbors (type: integer, default: 5)
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer, W. P. (2002). Smote: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16:321-357.
Gower, J. C. (1971). A general coefficient of similarity and some of its properties. Biometrics 27(4):857-871. (For the distance metric used)
smotenc() for direct implementation
Other Steps for over-sampling:
step_adasyn(),
step_bsmote(),
step_kmeans_smote(),
step_rose(),
step_smogn(),
step_smote(),
step_smoten(),
step_svmsmote(),
step_upsample()
library(recipes) library(modeldata) data(hpc_data) orig <- count(hpc_data, class, name = "orig") orig up_rec <- recipe(class ~ ., data = hpc_data) |> # Bring the minority levels up to about 1000 each # 1000/2211 is approx 0.4523 step_smotenc(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_data) |> count(class, name = "baked") baked # Note that if the original data contained more rows than the # target n (= ratio * majority_n), the data are left alone: orig |> left_join(training, by = "class") |> left_join(baked, by = "class")library(recipes) library(modeldata) data(hpc_data) orig <- count(hpc_data, class, name = "orig") orig up_rec <- recipe(class ~ ., data = hpc_data) |> # Bring the minority levels up to about 1000 each # 1000/2211 is approx 0.4523 step_smotenc(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_data) |> count(class, name = "baked") baked # Note that if the original data contained more rows than the # target n (= ratio * majority_n), the data are left alone: orig |> left_join(training, by = "class") |> left_join(baked, by = "class")
step_svmsmote() creates a specification of a recipe step that generate
new examples of the minority class near the decision boundary using the
support vectors of a fitted support vector machine.
step_svmsmote( recipe, ..., role = NA, trained = FALSE, column = NULL, over_ratio = 1, neighbors = 5, distance = "euclidean", m_neighbors = NULL, out_step = 0.5, indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("svmsmote") )step_svmsmote( recipe, ..., role = NA, trained = FALSE, column = NULL, over_ratio = 1, neighbors = 5, distance = "euclidean", m_neighbors = NULL, out_step = 0.5, indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("svmsmote") )
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 |
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 |
over_ratio |
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 |
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
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
m_neighbors |
An integer or |
out_step |
A number. Step size used when extrapolating new examples away from safe support vectors. Defaults to 0.5. |
indicator_column |
A single string or |
skip |
A logical. Should the step be skipped when the recipe is baked by
|
seed |
An integer that will be used as the seed when applied. |
id |
A character string that is unique to this step to identify it. |
SVM-SMOTE (Support Vector Machine SMOTE) works the same way as SMOTE, except that instead of generating points around every point of the minority class, it focuses generation near the decision boundary. A support vector machine is fitted to the data and the support vectors that belong to the minority class are used as the base points for generating new examples.
For each minority support vector its nearest neighbors among all classes are calculated. If all of the neighbors come from a different class the support vector is labeled noise and is discarded. If more than half of the neighbors come from a different class the support vector is labeled "danger" and new points are interpolated between it and its minority-class neighbors. The remaining support vectors are considered to be in a safe region and new points are extrapolated away from their minority-class neighbors.
The number of neighbors used for this labeling is controlled by
m_neighbors, and how far the extrapolated points are placed is controlled by
out_step.
The support vector machine is always fitted with kernlab::ksvm() using a
radial basis function kernel (kernel = "rbfdot") and a cost of C = 1,
matching the reference implementations of the method. These are not exposed
as arguments because the fitted model is only used to identify which minority
observations are support vectors, and not for prediction, so the sampling
results are relatively insensitive to them.
SMOTE generates new examples of the minority class using nearest neighbors
of these cases. For each existing minority class example, new examples are
created by interpolating between the example and its nearest neighbors. The
number of nearest neighbors used is controlled by the number of neighbors
argument (k in smote(), neighbors in step_smote()), and the number
of new examples generated is controlled by over_ratio.
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.
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.
Each minority class must have at least neighbors + 1 observations to
perform the SVM-SMOTE algorithm.
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: 5)
m_neighbors: # Nearest Neighbors (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.
Nguyen, H. M., Cooper, E. W., and Kamei, K. (2011). Borderline over-sampling for imbalanced data classification. International Journal of Knowledge Engineering and Soft Data Paradigms, 3(1), 4-21.
svmsmote() for direct implementation
Other Steps for over-sampling:
step_adasyn(),
step_bsmote(),
step_kmeans_smote(),
step_rose(),
step_smogn(),
step_smote(),
step_smoten(),
step_smotenc(),
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_svmsmote(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 SMOTE") recipe(class ~ x + y, data = circle_example) |> step_svmsmote(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With SVM-SMOTE")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_svmsmote(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 SMOTE") recipe(class ~ x + y, data = circle_example) |> step_svmsmote(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With SVM-SMOTE")
step_tomek() creates a specification of a recipe step that removes
majority class instances of tomek links.
step_tomek( recipe, ..., role = NA, trained = FALSE, column = NULL, distance = "euclidean", skip = TRUE, seed = sample.int(10^5, 1), distance_with = recipes::all_predictors(), id = rand_id("tomek") )step_tomek( recipe, ..., role = NA, trained = FALSE, column = NULL, distance = "euclidean", skip = TRUE, seed = sample.int(10^5, 1), distance_with = recipes::all_predictors(), id = rand_id("tomek") )
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 |
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 |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
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
|
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
|
id |
A character string that is unique to this step to identify it. |
A Tomek link is a pair of points from different classes that are each other's nearest neighbors. Such pairs sit on or very near the decision boundary and are considered noise or borderline cases. The algorithm identifies all Tomek links and removes the majority class instance from each pair, cleaning the class boundary without discarding non-boundary majority examples. Because only boundary points are removed, this typically discards far fewer observations than other under-sampling methods.
All variables selected by distance_with must be numeric with no missing
data.
All columns in the data are sampled and returned by recipes::juice()
and recipes::bake().
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.
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.
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
The underlying operation does not allow for case weights. Supplying data with a case weights column to this step results in an error.
Tomek. Two modifications of cnn. IEEE Trans. Syst. Man Cybern., 6:769-772, 1976.
tomek() for direct implementation
step_smote(), which is commonly composed before step_tomek() to clean
the ambiguous points that over-sampling creates near the class boundary
(the equivalent of imbalanced-learn's SMOTETomek).
Other Steps for under-sampling:
step_cluster_centroids(),
step_cnn(),
step_downsample(),
step_enn(),
step_instance_hardness(),
step_ncl(),
step_nearmiss(),
step_oss()
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) |> step_tomek(class) |> 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 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 Tomek") + xlim(c(1, 15)) + ylim(c(1, 15)) recipe(class ~ x + y, data = circle_example) |> step_tomek(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With Tomek") + xlim(c(1, 15)) + ylim(c(1, 15))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) |> step_tomek(class) |> 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 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 Tomek") + xlim(c(1, 15)) + ylim(c(1, 15)) recipe(class ~ x + y, data = circle_example) |> step_tomek(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_point() + labs(title = "With Tomek") + xlim(c(1, 15)) + ylim(c(1, 15))
step_upsample() creates a specification of a recipe step that will
replicate rows of a data set to make the occurrence of levels in a specific
factor level equal.
step_upsample( recipe, ..., over_ratio = 1, ratio = deprecated(), role = NA, trained = FALSE, column = NULL, target = NA, indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("upsample") )step_upsample( recipe, ..., over_ratio = 1, ratio = deprecated(), role = NA, trained = FALSE, column = NULL, target = NA, indicator_column = NULL, skip = TRUE, seed = sample.int(10^5, 1), id = rand_id("upsample") )
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 |
over_ratio |
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 |
ratio |
Deprecated argument; same as |
role |
For new variables created by this step, what analysis role
should they be assigned? Only used when |
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 |
target |
A named numeric vector giving the number of rows to sample
each level up to. This should not be set by the user and will be
populated by |
indicator_column |
A single string or |
skip |
A logical. Should the step be skipped when the recipe is baked by
|
seed |
An integer that will be used as the seed when applied. |
id |
A character string that is unique to this step to identify it. |
Up-sampling is intended to be performed on the training set
alone. For this reason, the default is skip = TRUE.
If there are missing values in the factor variable that is used to define the sampling, missing data are selected at random in the same way that the other factor levels are sampled. Missing values are not used to determine the amount of data in the majority level (see example below).
For any data with factor levels occurring with the same frequency as the majority level, all data will be retained.
All columns in the data are sampled and returned by recipes::juice()
and recipes::bake().
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.
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 1 tuning parameters:
over_ratio: Over-Sampling Ratio (type: double, default: 1)
This step performs an unsupervised operation that can utilize case weights.
To use them, see the documentation in recipes::case_weights and the examples on
tidymodels.org.
Other Steps for over-sampling:
step_adasyn(),
step_bsmote(),
step_kmeans_smote(),
step_rose(),
step_smogn(),
step_smote(),
step_smoten(),
step_smotenc(),
step_svmsmote()
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_upsample(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 # Note that if the original data contained more rows than the # target n (= ratio * majority_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 upsample") recipe(class ~ x + y, data = circle_example) |> step_upsample(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_jitter(width = 0.1, height = 0.1) + labs(title = "With upsample (with jittering)")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_upsample(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 # Note that if the original data contained more rows than the # target n (= ratio * majority_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 upsample") recipe(class ~ x + y, data = circle_example) |> step_upsample(class) |> prep() |> bake(new_data = NULL) |> ggplot(aes(x, y, color = class)) + geom_jitter(width = 0.1, height = 0.1) + labs(title = "With upsample (with jittering)")
SVM-SMOTE generates new examples of the minority class near the decision boundary, using the support vectors of a fitted SVM to decide where to place synthetic examples.
svmsmote( df, var, k = 5, over_ratio = 1, distance = "euclidean", m_neighbors = NULL, out_step = 0.5 )svmsmote( df, var, k = 5, over_ratio = 1, distance = "euclidean", m_neighbors = NULL, out_step = 0.5 )
df |
data.frame or tibble. Must have 1 factor variable and remaining numeric variables. |
var |
Character, name of variable containing factor variable. |
k |
An integer. Number of nearest neighbor that are used to generate the new examples of the minority class. |
over_ratio |
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 |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
The probability divergences are meaningful for compositional predictors such as proportions or counts normalized per observation, and are generally not appropriate for standardized predictors. |
m_neighbors |
An integer or |
out_step |
A number. Step size used when extrapolating new examples away from safe support vectors. Defaults to 0.5. |
SVM-SMOTE (Support Vector Machine SMOTE) works the same way as SMOTE, except that instead of generating points around every point of the minority class, it focuses generation near the decision boundary. A support vector machine is fitted to the data and the support vectors that belong to the minority class are used as the base points for generating new examples.
For each minority support vector its nearest neighbors among all classes are calculated. If all of the neighbors come from a different class the support vector is labeled noise and is discarded. If more than half of the neighbors come from a different class the support vector is labeled "danger" and new points are interpolated between it and its minority-class neighbors. The remaining support vectors are considered to be in a safe region and new points are extrapolated away from their minority-class neighbors.
The number of neighbors used for this labeling is controlled by
m_neighbors, and how far the extrapolated points are placed is controlled by
out_step.
The support vector machine is always fitted with kernlab::ksvm() using a
radial basis function kernel (kernel = "rbfdot") and a cost of C = 1,
matching the reference implementations of the method. These are not exposed
as arguments because the fitted model is only used to identify which minority
observations are support vectors, and not for prediction, so the sampling
results are relatively insensitive to them.
SMOTE generates new examples of the minority class using nearest neighbors
of these cases. For each existing minority class example, new examples are
created by interpolating between the example and its nearest neighbors. The
number of nearest neighbors used is controlled by the number of neighbors
argument (k in smote(), neighbors in step_smote()), and the number
of new examples generated is controlled by over_ratio.
All columns used in this function must be numeric with no missing data.
A data.frame or tibble, depending on type of df.
Nguyen, H. M., Cooper, E. W., and Kamei, K. (2011). Borderline over-sampling for imbalanced data classification. International Journal of Knowledge Engineering and Soft Data Paradigms, 3(1), 4-21.
step_svmsmote() for step function of this method
Other Direct Implementations:
adasyn(),
bsmote(),
cluster_centroids(),
cnn(),
enn(),
instance_hardness(),
kmeans_smote(),
ncl(),
nearmiss(),
oss(),
rose(),
smogn(),
smote(),
smoten(),
smotenc(),
tomek()
circle_numeric <- circle_example[, c("x", "y", "class")] res <- svmsmote(circle_numeric, var = "class") res <- svmsmote(circle_numeric, var = "class", k = 10) res <- svmsmote(circle_numeric, var = "class", over_ratio = 0.8) res <- svmsmote(circle_numeric, var = "class", distance = "manhattan") res <- svmsmote(circle_numeric, var = "class", m_neighbors = 20)circle_numeric <- circle_example[, c("x", "y", "class")] res <- svmsmote(circle_numeric, var = "class") res <- svmsmote(circle_numeric, var = "class", k = 10) res <- svmsmote(circle_numeric, var = "class", over_ratio = 0.8) res <- svmsmote(circle_numeric, var = "class", distance = "manhattan") res <- svmsmote(circle_numeric, var = "class", m_neighbors = 20)
Removes the majority class member of each pair of observations that form a Tomek link.
tomek(df, var, distance = "euclidean")tomek(df, var, distance = "euclidean")
df |
data.frame or tibble. Must have 1 factor variable and remaining numeric variables. |
var |
Character, name of variable containing factor variable. |
distance |
A character string specifying the distance metric used for
nearest neighbor calculations, defaulting to
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 Tomek link is a pair of points from different classes that are each other's nearest neighbors. Such pairs sit on or very near the decision boundary and are considered noise or borderline cases. The algorithm identifies all Tomek links and removes the majority class instance from each pair, cleaning the class boundary without discarding non-boundary majority examples. Because only boundary points are removed, this typically discards far fewer observations than other under-sampling methods.
All columns used in this function must be numeric with no missing data.
A data.frame or tibble, depending on type of df.
Tomek. Two modifications of cnn. IEEE Trans. Syst. Man Cybern., 6:769-772, 1976.
step_tomek() for step function of this method
Other Direct Implementations:
adasyn(),
bsmote(),
cluster_centroids(),
cnn(),
enn(),
instance_hardness(),
kmeans_smote(),
ncl(),
nearmiss(),
oss(),
rose(),
smogn(),
smote(),
smoten(),
smotenc(),
svmsmote()
circle_numeric <- circle_example[, c("x", "y", "class")] res <- tomek(circle_numeric, var = "class") res <- tomek(circle_numeric, var = "class", distance = "manhattan")circle_numeric <- circle_example[, c("x", "y", "class")] res <- tomek(circle_numeric, var = "class") res <- tomek(circle_numeric, var = "class", distance = "manhattan")