R/corr.R

In recipes: Preprocessing and Feature Engineering Steps for Modeling

#' High Correlation Filter
#'
#' `step_corr()` creates a *specification* of a recipe step that will
#' potentially remove variables that have large absolute correlations with other
#' variables.
#'
#' @inheritParams step_center
#' @param threshold A value for the threshold of absolute
#'  correlation values. The step will try to remove the minimum
#'  number of columns so that all the resulting absolute
#'  correlations are less than this value.
#' @param use A character string for the `use` argument to
#'  the [stats::cor()] function.
#' @param method A character string for the `method` argument
#'  to the [stats::cor()] function.
#' @param removals A character string that contains the names of
#'  columns that should be removed. These values are not determined
#'  until [prep()] is called.
#' @template step-return
#' @template filter-steps
#' @author Original R code for filtering algorithm by Dong Li,
#'  modified by Max Kuhn. Contributions by Reynald Lescarbeau (for
#'  original in `caret` package). Max Kuhn for the `step`
#'  function.
#' @family variable filter steps
#' @export
#'
#' @details This step attempts to remove variables to keep the
#'  largest absolute correlation between the variables less than
#'  `threshold`.
#'
#' When a column has a single unique value, that column will be
#'  excluded from the correlation analysis. Also, if the data set
#'  has sporadic missing values (and an inappropriate value of `use`
#'  is chosen), some columns will also be excluded from the filter.
#'
#' The arguments `use` and `method` don't take effect if case weights
#' are used in the recipe.
#'
#' # Tidying
#'
#' When you [`tidy()`][tidy.recipe()] this step, a tibble with column
#' `terms` (the columns that will be removed) is returned.
#'
#' ```{r, echo = FALSE, results="asis"}
#' step <- "step_corr"
#' result <- knitr::knit_child("man/rmd/tunable-args.Rmd")
#' cat(result)
#' ```
#'
#' @template case-weights-unsupervised
#'
#' @examplesIf rlang::is_installed("modeldata")
#' data(biomass, package = "modeldata")
#'
#' set.seed(3535)
#' biomass$duplicate <- biomass$carbon + rnorm(nrow(biomass))
#'
#' biomass_tr <- biomass[biomass$dataset == "Training", ]
#' biomass_te <- biomass[biomass$dataset == "Testing", ]
#'
#' rec <- recipe(
#'   HHV ~ carbon + hydrogen + oxygen + nitrogen + sulfur + duplicate,
#'   data = biomass_tr
#' )
#'
#' corr_filter <- rec %>%
#'   step_corr(all_numeric_predictors(), threshold = .5)
#'
#' filter_obj <- prep(corr_filter, training = biomass_tr)
#'
#' filtered_te <- bake(filter_obj, biomass_te)
#' round(abs(cor(biomass_tr[, c(3:7, 9)])), 2)
#' round(abs(cor(filtered_te)), 2)
#'
#' tidy(corr_filter, number = 1)
#' tidy(filter_obj, number = 1)
step_corr <- function(recipe,
                      ...,
                      role = NA,
                      trained = FALSE,
                      threshold = 0.9,
                      use = "pairwise.complete.obs",
                      method = "pearson",
                      removals = NULL,
                      skip = FALSE,
                      id = rand_id("corr")) {
  add_step(
    recipe,
    step_corr_new(
      terms = enquos(...),
      role = role,
      trained = trained,
      threshold = threshold,
      use = use,
      method = method,
      removals = removals,
      skip = skip,
      id = id,
      case_weights = NULL
    )
  )
}

step_corr_new <-
  function(terms, role, trained, threshold, use, method,
           removals, skip, id, case_weights) {
    step(
      subclass = "corr",
      terms = terms,
      role = role,
      trained = trained,
      threshold = threshold,
      use = use,
      method = method,
      removals = removals,
      skip = skip,
      id = id,
      case_weights = case_weights
    )
  }

#' @export
prep.step_corr <- function(x, training, info = NULL, ...) {
  col_names <- recipes_eval_select(x$terms, training, info)
  check_type(training[, col_names], types = c("double", "integer"))

  wts <- get_case_weights(info, training)
  were_weights_used <- are_weights_used(wts, unsupervised = TRUE)
  if (isFALSE(were_weights_used)) {
    wts <- NULL
  }

  if (length(col_names) > 1) {
    filter <- corr_filter(
      x = training[, col_names],
      wts = wts,
      cutoff = x$threshold,
      use = x$use,
      method = x$method
    )
  } else {
    filter <- character(0)
  }

  step_corr_new(
    terms = x$terms,
    role = x$role,
    trained = TRUE,
    threshold = x$threshold,
    use = x$use,
    method = x$method,
    removals = filter,
    skip = x$skip,
    id = x$id,
    case_weights = were_weights_used
  )
}

#' @export
bake.step_corr <- function(object, new_data, ...) {
  new_data <- recipes_remove_cols(new_data, object)
  new_data
}

print.step_corr <-
  function(x, width = max(20, options()$width - 36), ...) {
    title <- "Correlation filter on "
    print_step(x$removals, x$terms, x$trained, title, width,
               case_weights = x$case_weights)
    invisible(x)
  }

corr_filter <-
  function(x,
           wts = NULL,
           cutoff = .90,
           use = "pairwise.complete.obs",
           method = "pearson") {
    x <- correlations(x, wts = wts, use = use, method = method)

    if (any(!complete.cases(x))) {
      all_na <- apply(x, 2, function(x) all(is.na(x)))
      if (sum(all_na) >= nrow(x) - 1) {
        rlang::warn("Too many correlations are `NA`; skipping correlation filter.")
        return(numeric(0))
      } else {
        na_cols <- which(all_na)
        if (length(na_cols) > 0) {
          x[na_cols, ] <- 0
          x[, na_cols] <- 0
          rlang::warn(
            paste0(
              "The correlation matrix has missing values. ",
              length(na_cols),
              " columns were excluded from the filter."
            )
          )
        }
      }
      if (any(is.na(x))) {
        rlang::warn(
          paste0(
            "The correlation matrix has sporadic missing values. ",
            "Some columns were excluded from the filter."
          )
        )
        x[is.na(x)] <- 0
      }
      diag(x) <- 1
    }
    averageCorr <- colMeans(abs(x))
    averageCorr <- as.numeric(as.factor(averageCorr))
    x[lower.tri(x, diag = TRUE)] <- NA
    combsAboveCutoff <- which(abs(x) > cutoff)

    colsToCheck <- ceiling(combsAboveCutoff / nrow(x))
    rowsToCheck <- combsAboveCutoff %% nrow(x)

    colsToDiscard <- averageCorr[colsToCheck] > averageCorr[rowsToCheck]
    rowsToDiscard <- !colsToDiscard

    deletecol <- c(colsToCheck[colsToDiscard], rowsToCheck[rowsToDiscard])
    deletecol <- unique(deletecol)
    if (length(deletecol) > 0) {
      deletecol <- colnames(x)[deletecol]
    }
    deletecol
  }

tidy_filter <- function(x, ...) {
  if (is_trained(x)) {
    res <- tibble(terms = unname(x$removals))
  } else {
    term_names <- sel2char(x$terms)
    res <- tibble(terms = term_names)
  }
  res$id <- x$id
  res
}

#' @rdname tidy.recipe
#' @export
tidy.step_corr <- tidy_filter

#' @export
tunable.step_corr <- function(x, ...) {
  tibble::tibble(
    name = "threshold",
    call_info = list(
      list(pkg = "dials", fun = "threshold")
    ),
    source = "recipe",
    component = "step_corr",
    component_id = x$id
  )
}