R/prob-classification_cost.R

In yardstick: Tidy Characterizations of Model Performance

#' Costs function for poor classification
#'
#' @description
#' `classification_cost()` calculates the cost of a poor prediction based on
#' user-defined costs. The costs are multiplied by the estimated class
#' probabilities and the mean cost is returned.
#'
#' @details
#' As an example, suppose that there are three classes: `"A"`, `"B"`, and `"C"`.
#' Suppose there is a truly `"A"` observation with class probabilities `A = 0.3
#' / B = 0.3 / C = 0.4`. Suppose that, when the true result is class `"A"`, the
#' costs for each class were `A = 0 / B = 5 / C = 10`, penalizing the
#' probability of incorrectly predicting `"C"` more than predicting `"B"`. The
#' cost for this prediction would be `0.3 * 0 + 0.3 * 5 + 0.4 * 10`. This
#' calculation is done for each sample and the individual costs are averaged.
#'
#' @family class probability metrics
#' @templateVar fn class_cost
#' @template return
#'
#' @inheritParams pr_auc
#'
#' @param costs A data frame with columns `"truth"`, `"estimate"`, and `"cost"`.
#'
#' `"truth"` and `"estimate"` should be character columns containing unique
#' combinations of the levels of the `truth` factor.
#'
#' `"costs"` should be a numeric column representing the cost that should
#' be applied when the `"estimate"` is predicted, but the true result is
#' `"truth"`.
#'
#' It is often the case that when `"truth" == "estimate"`, the cost is zero
#' (no penalty for correct predictions).
#'
#' If any combinations of the levels of `truth` are missing, their costs are
#' assumed to be zero.
#'
#' If `NULL`, equal costs are used, applying a cost of `0` to correct
#' predictions, and a cost of `1` to incorrect predictions.
#'
#' @author Max Kuhn
#'
#' @export
#' @examples
#' library(dplyr)
#'
#' # ---------------------------------------------------------------------------
#' # Two class example
#' data(two_class_example)
#'
#' # Assuming `Class1` is our "event", this penalizes false positives heavily
#' costs1 <- tribble(
#'   ~truth,   ~estimate, ~cost,
#'   "Class1", "Class2",  1,
#'   "Class2", "Class1",  2
#' )
#'
#' # Assuming `Class1` is our "event", this penalizes false negatives heavily
#' costs2 <- tribble(
#'   ~truth,   ~estimate, ~cost,
#'   "Class1", "Class2",  2,
#'   "Class2", "Class1",  1
#' )
#'
#' classification_cost(two_class_example, truth, Class1, costs = costs1)
#'
#' classification_cost(two_class_example, truth, Class1, costs = costs2)
#'
#' # ---------------------------------------------------------------------------
#' # Multiclass
#' data(hpc_cv)
#'
#' # Define cost matrix from Kuhn and Johnson (2013)
#' hpc_costs <- tribble(
#'   ~estimate, ~truth, ~cost,
#'   "VF",      "VF",    0,
#'   "VF",      "F",     1,
#'   "VF",      "M",     5,
#'   "VF",      "L",    10,
#'   "F",       "VF",    1,
#'   "F",       "F",     0,
#'   "F",       "M",     5,
#'   "F",       "L",     5,
#'   "M",       "VF",    1,
#'   "M",       "F",     1,
#'   "M",       "M",     0,
#'   "M",       "L",     1,
#'   "L",       "VF",    1,
#'   "L",       "F",     1,
#'   "L",       "M",     1,
#'   "L",       "L",     0
#' )
#'
#' # You can use the col1:colN tidyselect syntax
#' hpc_cv %>%
#'   filter(Resample == "Fold01") %>%
#'   classification_cost(obs, VF:L, costs = hpc_costs)
#'
#' # Groups are respected
#' hpc_cv %>%
#'   group_by(Resample) %>%
#'   classification_cost(obs, VF:L, costs = hpc_costs)
classification_cost <- function(data, ...) {
  UseMethod("classification_cost")
}
classification_cost <- new_prob_metric(
  fn = classification_cost,
  direction = "minimize"
)

#' @rdname classification_cost
#' @export
classification_cost.data.frame <- function(data,
                                           truth,
                                           ...,
                                           costs = NULL,
                                           na_rm = TRUE,
                                           event_level = yardstick_event_level(),
                                           case_weights = NULL) {
  prob_metric_summarizer(
    name = "classification_cost",
    fn = classification_cost_vec,
    data = data,
    truth = !!enquo(truth),
    ...,
    na_rm = na_rm,
    event_level = event_level,
    case_weights = !!enquo(case_weights),
    # Extra argument for classification_cost_impl()
    fn_options = list(costs = costs)
  )
}

#' @rdname classification_cost
#' @export
classification_cost_vec <- function(truth,
                                    estimate,
                                    costs = NULL,
                                    na_rm = TRUE,
                                    event_level = yardstick_event_level(),
                                    case_weights = NULL,
                                    ...) {
  abort_if_class_pred(truth)

  estimator <- finalize_estimator(truth, metric_class = "classification_cost")

  check_prob_metric(truth, estimate, case_weights, estimator)

  if (na_rm) {
    result <- yardstick_remove_missing(truth, estimate, case_weights)

    truth <- result$truth
    estimate <- result$estimate
    case_weights <- result$case_weights
  } else if (yardstick_any_missing(truth, estimate, case_weights)) {
    return(NA_real_)
  }

  classification_cost_estimator_impl(
    truth = truth,
    estimate = estimate,
    costs = costs,
    estimator = estimator,
    event_level = event_level,
    case_weights = case_weights
  )
}

classification_cost_estimator_impl <- function(truth,
                                               estimate,
                                               costs,
                                               estimator,
                                               event_level,
                                               case_weights) {
  if (is_binary(estimator)) {
    classification_cost_binary(truth, estimate, costs, event_level, case_weights)
  } else {
    classification_cost_multiclass(truth, estimate, costs, case_weights)
  }
}

classification_cost_binary <- function(truth,
                                       estimate,
                                       costs,
                                       event_level,
                                       case_weights) {
  if (is_event_first(event_level)) {
    level1 <- estimate
    level2 <- 1 - estimate
  } else {
    level1 <- 1 - estimate
    level2 <- estimate
  }

  estimate <- c(level1, level2)
  estimate <- matrix(estimate, ncol = 2)

  classification_cost_multiclass(truth, estimate, costs, case_weights)
}

classification_cost_multiclass <- function(truth, estimate, costs, case_weights) {
  levels <- levels(truth)

  costs <- validate_costs(costs, levels)
  costs <- pivot_costs(costs, levels)
  costs <- recycle_costs(costs, truth)

  costs <- costs[levels]
  costs <- as.matrix(costs)

  out <- estimate * costs
  out <- rowSums(out)
  out <- yardstick_mean(out, case_weights = case_weights)

  out
}

validate_costs <- function(costs, levels) {
  if (is.null(costs)) {
    costs <- generate_equal_cost_grid(levels)
    return(costs)
  }

  if (!is.data.frame(costs)) {
    abort("`costs` must be `NULL` or a data.frame.")
  }

  columns <- names(costs)
  if (length(columns) != 3L) {
    abort("`costs` must be a data.frame with 3 columns.")
  }

  ok <- identical(sort(columns), sort(c("truth", "estimate", "cost")))
  if (!ok) {
    abort("`costs` must have columns: 'truth', 'estimate', and 'cost'.")
  }

  if (is.factor(costs$truth)) {
    costs$truth <- as.character(costs$truth)
  }
  if (!is.character(costs$truth)) {
    abort("`costs$truth` must be a character or factor column.")
  }

  if (is.factor(costs$estimate)) {
    costs$estimate <- as.character(costs$estimate)
  }
  if (!is.character(costs$estimate)) {
    abort("`costs$estimate` must be a character or factor column.")
  }

  if (!is.numeric(costs$cost)) {
    abort("`costs$cost` must be a numeric column.")
  }

  ok <- all(costs$truth %in% levels)
  if (is_false(ok)) {
    levels <- quote_and_collapse(levels)
    abort(paste0("`costs$truth` can only contain ", levels, "."))
  }

  ok <- all(costs$estimate %in% levels)
  if (is_false(ok)) {
    levels <- quote_and_collapse(levels)
    abort(paste0("`costs$estimate` can only contain ", levels, "."))
  }

  pairs <- costs[c("truth", "estimate")]
  cost <- costs$cost

  not_ok <- vec_duplicate_any(pairs)
  if (not_ok) {
    abort("`costs` cannot have duplicate 'truth' / 'estimate' combinations.")
  }

  out <- generate_all_level_combinations(levels)

  # Detect user specified cost locations
  locs <- vec_match(pairs, out)

  # Default to no cost
  out$cost <- 0

  # Update to user specified cost
  out$cost[locs] <- cost

  out
}

# - 0 cost for correct predictions
# - 1 cost for incorrect predictions
generate_equal_cost_grid <- function(levels) {
  costs <- generate_all_level_combinations(levels)
  costs$cost <- ifelse(costs$truth == costs$estimate, yes = 0, no = 1)
  costs
}

generate_all_level_combinations <- function(levels) {
  # `expand.grid()` expands first column fastest,
  # but we want first column slowest so we reverse the columns
  grid <- expand.grid(estimate = levels, truth = levels)
  grid <- dplyr::as_tibble(grid)
  grid <- grid[c("truth", "estimate")]
  grid
}

pivot_costs <- function(costs, levels) {
  # Must be a data frame, not a tibble, for `reshape()` to work
  costs <- as.data.frame(costs)

  # tidyr::pivot_wider(costs, truth, names_from = estimate, values_from = cost)
  costs <- stats::reshape(
    costs,
    v.names = "cost",
    idvar = "truth",
    timevar = "estimate",
    direction = "wide",
    sep = "."
  )

  # Ensure column ordering matches `truth` level ordering
  columns <- paste0("cost.", levels)
  costs <- costs[c("truth", columns)]

  names(costs) <- c("truth", levels)

  costs <- dplyr::as_tibble(costs)

  costs
}

recycle_costs <- function(costs, truth) {
  levels <- levels(truth)

  # Expand `costs` to equal size of `truth`, matching the `truth` values
  needles <- truth
  haystack <- factor(costs$truth, levels = levels)
  locs <- vec_match(needles, haystack)

  costs <- vec_slice(costs, locs)

  costs
}