R/smote.R

In themis: Extra Recipes Steps for Dealing with Unbalanced Data

#' Apply SMOTE Algorithm
#'
#' `step_smote()` creates a *specification* of a recipe step that generate new
#' examples of the minority class using nearest neighbors of these cases.
#'
#' @inheritParams recipes::step_center
#' @inheritParams step_upsample
#' @param ... One or more selector functions to choose which
#'  variable is used to sample the data. See [recipes::selections]
#'  for more details. The selection should result in _single
#'  factor variable_. For the `tidy` method, these are not
#'  currently used.
#' @param column A character string of the variable name that will
#'  be populated (eventually) by the `...` selectors.
#' @param neighbors An integer. Number of nearest neighbor that are used
#'  to generate the new examples of the minority class.
#' @param distance A character string specifying the distance metric used for
#'  nearest neighbor calculations, defaulting to `"euclidean"`. The available
#'  metrics fall into three groups.
#'
#'  `"euclidean"`, `"cosine"`, and `"mahalanobis"` use approximate nearest
#'  neighbors via the RANN package and scale well to large datasets.
#'
#'  `"squared_chord"`, `"matusita"`, `"hellinger"`, and `"bhattacharyya"` are
#'  probability-divergence measures that treat each row as a distribution over
#'  the predictors, so they require non-negative values. `"hellinger"` and
#'  `"bhattacharyya"` further require each row to sum to 1. All four also use
#'  the RANN package and scale well to large datasets.
#'
#'  `"manhattan"`, `"chebyshev"`, `"canberra"`, `"soergel"`, `"lorentzian"`,
#'  `"jeffreys"`, `"topsoe"`, `"jensen-shannon"`, `"jensen_difference"`,
#'  `"taneja"`, and `"kumar-johnson"` compute an exact all-pairs distance
#'  matrix. This takes time and memory proportional to the square of the number
#'  of observations in a class, so these are best suited to smaller datasets.
#'  Everything from `"canberra"` onwards is a probability divergence requiring
#'  non-negative values, is provided by the philentropy package (which must be
#'  installed separately), and in the case of `"jeffreys"`, `"taneja"`, and
#'  `"kumar-johnson"` requires strictly positive values, since those divide by
#'  individual predictor values.
#'
#'  The probability divergences are meaningful for compositional predictors such
#'  as proportions or counts normalized per observation, and are generally not
#'  appropriate for standardized predictors.
#' @param seed An integer that will be used as the seed when applied.
#' @return 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.
#'
#' @template details-smote
#'
#' @details
#' 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.
#'
#' # Minimum observations
#'
#' Each minority class must have at least `neighbors + 1` observations to
#' perform the SMOTE algorithm.
#'
#' # Tidying
#'
#' When you [`tidy()`][recipes::tidy.recipe()] this step, a tibble is returned with
#'  columns `terms` and `id`:
#'
#' \describe{
#'   \item{terms}{character, the selectors or variables selected}
#'   \item{id}{character, id of this step}
#' }
#'
#' ```{r, echo = FALSE, results="asis"}
#' step <- "step_smote"
#' result <- knitr::knit_child("man/rmd/tunable-args.Rmd")
#' cat(result)
#' ```
#'
#' @template case-weights-not-supported
#'
#' @references 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.
#'
#' @seealso [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`).
#' @family Steps for over-sampling
#'
#' @export
#' @examplesIf rlang::is_installed("modeldata")
#' 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_smote <-
  function(
    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")
  ) {
    check_number_whole(seed)
    check_string(indicator_column, allow_null = TRUE, allow_empty = FALSE)
    check_distance_arg(distance)

    add_step(
      recipe,
      step_smote_new(
        terms = enquos(...),
        role = role,
        trained = trained,
        column = column,
        over_ratio = over_ratio,
        neighbors = neighbors,
        distance = distance,
        predictors = NULL,
        indicator_column = indicator_column,
        skip = skip,
        seed = seed,
        id = id
      )
    )
  }

step_smote_new <-
  function(
    terms,
    role,
    trained,
    column,
    over_ratio,
    neighbors,
    distance,
    predictors,
    indicator_column,
    skip,
    seed,
    id
  ) {
    step(
      subclass = "smote",
      terms = terms,
      role = role,
      trained = trained,
      column = column,
      over_ratio = over_ratio,
      neighbors = neighbors,
      distance = distance,
      predictors = predictors,
      indicator_column = indicator_column,
      skip = skip,
      seed = seed,
      id = id
    )
  }

#' @export
prep.step_smote <- function(x, training, info = NULL, ...) {
  x <- fill_new_args(x, list(distance = "euclidean"))

  col_name <- recipes_eval_select(x$terms, training, info)

  check_ratio(x$over_ratio, arg = "over_ratio")
  check_number_whole(x$neighbors, arg = "neighbors", min = 1)

  check_1_selected(col_name)
  check_column_factor(training, col_name)
  warn_unused_levels(training, col_name)
  check_ratio_column(x$over_ratio, training, col_name, arg = "over_ratio")

  recipes::check_name(
    tibble(x = logical(0)),
    training,
    x,
    newname = x$indicator_column
  )

  predictors <- setdiff(recipes::recipes_names_predictors(info), col_name)

  check_type(training[, predictors], types = c("double", "integer"))
  check_na(select(training, all_of(c(col_name, predictors))))

  step_smote_new(
    terms = x$terms,
    role = x$role,
    trained = TRUE,
    column = col_name,
    over_ratio = x$over_ratio,
    neighbors = x$neighbors,
    distance = x$distance,
    predictors = predictors,
    indicator_column = x$indicator_column,
    skip = x$skip,
    seed = x$seed,
    id = x$id
  )
}

#' @export
bake.step_smote <- function(object, new_data, ...) {
  object <- fill_new_args(object, list(distance = "euclidean"))

  col_names <- unique(c(object$predictors, object$column))
  check_new_data(col_names, object, new_data)

  if (length(object$column) == 0L) {
    # Empty selection
    return(new_data)
  }

  if (nrow(new_data) <= 1) {
    return(new_data)
  }

  n_orig <- nrow(new_data)
  new_data <- as.data.frame(new_data)

  predictor_data <- new_data[, col_names]

  # smote with seed for reproducibility
  with_seed(
    seed = object$seed,
    code = {
      synthetic_data <- smote_impl(
        predictor_data,
        object$column,
        k = object$neighbors,
        over_ratio = object$over_ratio,
        distance = object$distance
      )
      synthetic_data <- as_tibble(synthetic_data)
    }
  )
  new_data <- na_splice(new_data, synthetic_data, object)

  new_data <- add_indicator_column(new_data, n_orig, object$indicator_column)

  new_data
}

#' @export
print.step_smote <-
  function(x, width = max(20, options()$width - 26), ...) {
    title <- "SMOTE based on "
    print_step(x$column, x$terms, x$trained, title, width)
    invisible(x)
  }

#' @rdname step_smote
#' @usage NULL
#' @export
tidy.step_smote <- function(x, ...) {
  if (is_trained(x)) {
    res <- tibble(terms = unname(x$column))
  } else {
    term_names <- sel2char(x$terms)
    res <- tibble(terms = unname(term_names))
  }
  res$id <- x$id
  res
}

#' @export
#' @rdname tunable_themis
tunable.step_smote <- function(x, ...) {
  tibble::tibble(
    name = c("over_ratio", "neighbors"),
    call_info = list(
      list(pkg = "dials", fun = "over_ratio"),
      list(pkg = "dials", fun = "neighbors", range = c(1, 10))
    ),
    source = "recipe",
    component = "step_smote",
    component_id = x$id
  ) |>
    drop_per_class_ratio(x$over_ratio)
}

#' @rdname required_pkgs.step
#' @export
required_pkgs.step_smote <- function(x, ...) {
  c("themis")
}