R/errorest-bcv.r

In sortinghat: sortinghat

#' Calculates the Bootstrap Cross-Validation (BCV) Error Rate Estimator for a
#' specified classifier given a data set.
#'
#' For a given data matrix and its corresponding vector of labels, we calculate
#' the bootstrap cross-validation (BCV) error rate from Fu, Carroll, and Wang
#' (2005) for a given classifier.
#'
#' To calculate the BCV error rate, we sample from the data with replacement to
#' obtain a bootstrapped training data set. We then compute a cross-validation
#' error rate with the given classifier (given in \code{train}) on the
#' bootstrapped training data set. We repeat this process \code{num_bootstraps}
#' times to obtain a set of bootstrapped cross-validation error rates. We report
#' the average of these error rates. The \code{\link{errorest_cv}} function is
#' used to compute the cross-validation (CV) error rate estimator for each
#' bootstrap iteration.
#'
#' Fu et al. (2005) note that the BCV method works well because it is a bagging
#' classification error. Furthermore, consider the leave-one-out (LOO) error
#' rate estimator. For small sample sizes, the data are sparse, so that the left
#' out observation has a high probability of being far in distance from the
#' remaining training data set. Hence, the LOO error rate estimator yields a
#' large variance for small data sets.
#'
#' Rather than partitioning the observations into folds, an alternative
#' convention is to specify the 'hold-out' size for each test data set. Note that
#' this convention is equivalent to the notion of folds. We allow the user to
#' specify either option with the \code{hold_out} and \code{num_folds} arguments.
#' The \code{num_folds} argument is the default option but is ignored if the
#' \code{hold_out} argument is specified (i.e. is not \code{NULL}).
#'
#' We expect that the first two arguments of the classifier function given in
#' \code{train} are \code{x} and \code{y}, corresponding to the data matrix and
#' the vector of their labels. Additional arguments can be passed to the
#' \code{train} function. The returned object should be a classifier that will
#' be passed to the function given in the \code{classify} argument.
#'
#' We stay with the usual R convention for the \code{classify} function. We
#' expect that this function takes two arguments: 1. an \code{object} argument
#' which contains the trained classifier returned from the function specified in
#' \code{train}; and 2. a \code{newdata} argument which contains a matrix of
#' observations to be classified -- the matrix should have rows corresponding to
#' the individual observations and columns corresponding to the features
#' (covariates).
#'
#' @export
#' @references Fu, W.J., Carroll, R.J., and Wang, S. (2005), "Estimating
#' misclassification error with small samples via bootstrap cross-validation,"
#' Bioinformatics, vol. 21, no. 9, pp. 1979-1986.
#' @param x a matrix of n observations (rows) and p features (columns)
#' @param y a vector of n class labels
#' @param train a function that builds the classifier. (See details.)
#' @param classify a function that classifies observations from the constructed
#' classifier from \code{train}. (See details.)
#' @param num_bootstraps the number of bootstrap replications
#' @param num_folds the number of cross-validation folds. Ignored if
#' \code{hold_out} is not \code{NULL}. See Details.
#' @param hold_out the hold-out size for cross-validation. See Details.
#' @param ... additional arguments passed to the function specified in
#' \code{train}.
#' @return the BCV error rate estimate
#' @examples
#' require('MASS')
#' iris_x <- data.matrix(iris[, -5])
#' iris_y <- iris[, 5]
#'
#' # Because the \code{classify} function returns multiples objects in a list,
#' # we provide a wrapper function that returns only the class labels.
#' lda_wrapper <- function(object, newdata) { predict(object, newdata)$class }
#' set.seed(42)
#' errorest_bcv(x = iris_x, y = iris_y, train = MASS:::lda,
#'              classify = lda_wrapper)
#' # Output: 0.02213333
#'
errorest_bcv <- function(x, y, train, classify, num_bootstraps = 50,
                         num_folds = 10, hold_out = NULL, ...) {
  x <- as.matrix(x)
  y <- as.factor(y)
  check_out <- check_arguments(x = x, y = y, train = train, classify = classify)

  # For each bootstrap replicate, we sample with replacement from the data set
  # and compute the cross-validation error rate from the bootstrapped data set.
  bcv_error_rates <- sapply(seq_len(num_bootstraps), function(b) {
    training <- sample(seq_along(y), replace = TRUE)
    errorest_cv(x = x[training, ], y = y[training], train = train,
                classify = classify, num_folds = num_folds,
                hold_out = hold_out, ...)
  })
  
  mean(bcv_error_rates)
}