Extra Recipes Steps for Dealing with Unbalanced Data

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 , BorderlineSMOTE 2005 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.

themis

themis contains extra steps for the recipes package for dealing with unbalanced data. The name themis is that of the ancient Greek god who is typically depicted with a balance.

themis handles imbalance in both classification and regression problems: alongside the many classification samplers, step_smogn() resamples an imbalanced numeric outcome. Its nearest-neighbor-based steps are also not limited to Euclidean distance, the distance argument supports euclidean, cosine, mahalanobis, manhattan, and chebyshev metrics.

Installation

You can install the released version of themis from CRAN with:

install.packages("themis")

Install the development version from GitHub with:

# install.packages("pak")
pak::pak("tidymodels/themis")

Example

Following is an example of using the SMOTE algorithm to deal with unbalanced data

library(recipes)
library(modeldata)
library(themis)

data("credit_data", package = "modeldata")

credit_data0 <- credit_data |>
  filter(!is.na(Job))

count(credit_data0, Job)
#>         Job    n
#> 1     fixed 2805
#> 2 freelance 1024
#> 3    others  171
#> 4   partime  452

ds_rec <- recipe(Job ~ Time + Age + Expenses, data = credit_data0) |>
  step_impute_mean(all_predictors()) |>
  step_smote(Job, over_ratio = 0.25) |>
  prep()

ds_rec |>
  bake(new_data = NULL) |>
  count(Job)
#> # A tibble: 4 × 2
#>   Job           n
#>   <fct>     <int>
#> 1 fixed      2805
#> 2 freelance  1024
#> 3 others      701
#> 4 partime     701

Methods

Below is some unbalanced data. Used for examples later.

example_data <- data.frame(class = letters[rep(1:5, 1:5 * 10)],
                           x = rnorm(150))

library(ggplot2)

example_data |>
  ggplot(aes(class)) +
  geom_bar()

Bar chart with 5 columns. class on the x-axis and count on the y-axis. Class a has height 10, b has 20, c has 30, d has 40, and e has 50.

Upsample / Over-sampling

The following methods all share the tuning parameter over_ratio, which is the ratio of the minority-to-majority frequencies.

name	function	Multi-class
Random minority over-sampling with replacement	`step_upsample()`	:heavy_check_mark:
Synthetic Minority Over-sampling Technique	`step_smote()`	:heavy_check_mark:
SMOTE for datasets with continuous and nominal features	`step_smotenc()`	:heavy_check_mark:
SMOTE for nominal features only	`step_smoten()`	:heavy_check_mark:
Borderline SMOTE-1	`step_bsmote(all_neighbors = FALSE)`	:heavy_check_mark:
Borderline SMOTE-2	`step_bsmote(all_neighbors = TRUE)`	:heavy_check_mark:
Support-vector SMOTE	`step_svmsmote()`	:heavy_check_mark:
KMeans-SMOTE	`step_kmeans_smote()`	:heavy_check_mark:
Adaptive synthetic sampling approach for imbalanced learning	`step_adasyn()`	:heavy_check_mark:
Generation of synthetic data by Randomly Over Sampling Examples	`step_rose()`

step_smogn() also over-samples, but for imbalanced regression rather than classification: it resamples a numeric outcome instead of a class, so it does not use over_ratio and is not shown in the table above.

By setting over_ratio = 1 you bring the number of samples of all minority classes equal to 100% of the majority class.

recipe(~., example_data) |>
  step_upsample(class, over_ratio = 1) |>
  prep() |>
  bake(new_data = NULL) |>
  ggplot(aes(class)) +
  geom_bar()

Bar chart with 5 columns. class on the x-axis and count on the y-axis. class a, b, c, d, and e all have a height of 50.

and by setting over_ratio = 0.5 we upsample any minority class with fewer samples than 50% of the majority up to have 50% of the majority.

recipe(~., example_data) |>
  step_upsample(class, over_ratio = 0.5) |>
  prep() |>
  bake(new_data = NULL) |>
  ggplot(aes(class)) +
  geom_bar()

Bar chart with 5 columns. class on the x-axis and count on the y-axis. Class a has height 25, b has 25, c has 30, d has 40, and e has 50.

Downsample / Under-sampling

Most of the following methods all share the tuning parameter under_ratio, which is the ratio of the majority-to-minority frequencies.

name	function	Multi-class	under_ratio
Random majority under-sampling with replacement	`step_downsample()`	:heavy_check_mark:	:heavy_check_mark:
NearMiss-1, NearMiss-2, and NearMiss-3	`step_nearmiss(version = )`	:heavy_check_mark:	:heavy_check_mark:
Instance hardness threshold	`step_instance_hardness()`	:heavy_check_mark:	:heavy_check_mark:
Cluster centroids	`step_cluster_centroids()`	:heavy_check_mark:	:heavy_check_mark:
Condensed nearest neighbors	`step_cnn()`	:heavy_check_mark:
Edited nearest neighbors	`step_enn()`	:heavy_check_mark:
Neighborhood cleaning rule	`step_ncl()`	:heavy_check_mark:
One-sided selection	`step_oss()`	:heavy_check_mark:
Extraction of majority-minority Tomek links	`step_tomek()`

By setting under_ratio = 1 you bring the number of samples of all majority classes equal to 100% of the minority class.

recipe(~., example_data) |>
  step_downsample(class, under_ratio = 1) |>
  prep() |>
  bake(new_data = NULL) |>
  ggplot(aes(class)) +
  geom_bar()

Bar chart with 5 columns. class on the x-axis and count on the y-axis. Class a, b, c, d, and e all have a height of 10.

and by setting under_ratio = 2 we downsample any majority class with more than 200% samples of the minority class down to have 200% samples of the minority.

recipe(~., example_data) |>
  step_downsample(class, under_ratio = 2) |>
  prep() |>
  bake(new_data = NULL) |>
  ggplot(aes(class)) +
  geom_bar()

Bar chart with 5 columns. class on the x-axis and count on the y-axis. Class a has height 10, b, c, d, and e have ha height of 20.

Contributing

This project is released with a Contributor Code of Conduct. By contributing to this project, you agree to abide by its terms.

For questions and discussions about tidymodels packages, modeling, and machine learning, join us on RStudio Community.
If you think you have encountered a bug, please submit an issue.
Either way, learn how to create and share a reprex (a minimal, reproducible example), to clearly communicate about your code.
Check out further details on contributing guidelines for tidymodels packages and how to get help.