Merge pull request #571 from bgreenwell/master

New deforest() function for removing trees from a fitted random forest

NAMESPACE

@@ -1,16 +1,19 @@
 # Generated by roxygen2: do not edit by hand
 
+S3method(deforest,ranger)
 S3method(importance,ranger)
 S3method(predict,ranger)
 S3method(predict,ranger.forest)
 S3method(predictions,ranger)
 S3method(predictions,ranger.prediction)
+S3method(print,deforest.ranger)
 S3method(print,ranger)
 S3method(print,ranger.forest)
 S3method(print,ranger.prediction)
 S3method(timepoints,ranger)
 S3method(timepoints,ranger.prediction)
 export(csrf)
+export(deforest)
 export(getTerminalNodeIDs)
 export(holdoutRF)
 export(importance)

R/deforest.R

@@ -0,0 +1,175 @@
+#' Deforesting a random forest
+#' 
+#' The main purpose of this function is to allow for post-processing of 
+#' ensembles via L2 regularized regression (i.e., the LASSO), as described in 
+#' Friedman and Popescu (2003). The basic idea is to use the LASSO to 
+#' post-process the predictions from the individual base learners in an ensemble 
+#' (i.e., decision trees) in the hopes of producing a much smaller model without 
+#' sacrificing much in the way of accuracy, and in some cases, improving it. 
+#' Friedman and Popescu (2003) describe conditions under which tree-based 
+#' ensembles, like random forest, can potentially benefit from such 
+#' post-processing (e.g., using shallower trees trained on much smaller samples 
+#' of the training data without replacement). However, the computational 
+#' benefits of such post-processing can only be realized if the base learners 
+#' "zeroed out" by the LASSO can actually be removed from the original ensemble, 
+#' hence the purpose of this function. A complete example using 
+#' \code{\link{ranger}} can be found at 
+#' \url{https://github.com/imbs-hl/ranger/issues/568}.
+#' 
+#' @param object A fitted random forest (e.g., a \code{\link{ranger}}
+#' object).
+#' 
+#' @param which.trees Vector giving the indices of the trees to remove.
+#' 
+#' @param warn Logical indicating whether or not to warn users that some of the 
+#' standard output of a typical \code{\link{ranger}} object or no longer 
+#' available after deforestation. Default is \code{TRUE}.
+#' 
+#' @param ... Additional (optional) arguments. (Currently ignored.)
+#' 
+#' @return An object of class \code{"deforest.ranger"}; essentially, a 
+#' \code{\link{ranger}} object with certain components replaced with 
+#' \code{NA}s (e.g., out-of-bag (OOB) predictions, variable importance scores 
+#' (if requested), and OOB-based error metrics).
+#' 
+#' @note This function is a generic and can be extended by other packages.
+#' 
+#' @references 
+#' Friedman, J. and Popescu, B. (2003). Importance sampled learning ensembles, 
+#' Technical report, Stanford University, Department of Statistics.
+#' \url{https://statweb.stanford.edu/~jhf/ftp/isle.pdf}.
+#' 
+#' @rdname deforest
+#' 
+#' @export
+#' 
+#' @author Brandon M. Greenwell
+#' 
+#' @examples
+#' ## Example of deforesting a random forest
+#' rfo <- ranger(Species ~ ., data = iris, probability = TRUE, num.trees = 100)
+#' dfo <- deforest(rfo, which.trees = c(1, 3, 5))
+#' dfo  # same as `rfo` but with trees 1, 3, and 5 removed
+#' 
+#' ## Sanity check
+#' preds.rfo <- predict(rfo, data = iris, predict.all = TRUE)$predictions
+#' preds.dfo <- predict(dfo, data = iris, predict.all = TRUE)$predictions
+#' identical(preds.rfo[, , -c(1, 3, 5)], y = preds.dfo)
+deforest <- function(object, which.trees = NULL, ...) {
+  UseMethod("deforest")
+}
+
+
+#' @rdname deforest
+#' 
+#' @export
+deforest.ranger <- function(object, which.trees = NULL, warn = TRUE, ...) {
+
+  # Warn users about `predictions` and `prediction.error` components
+  if (isTRUE(warn)) {
+    warning("Many of the components of a typical \"ranger\" object are ",
+            "not available after deforestation and are instead replaced with ",
+            "`NA` (e.g., out-of-bag (OOB) predictions, variable importance ",
+            "scores (if requested), and OOB-based error metrics).", 
+            call. = FALSE)
+  }
+
+  # "Remove trees" by removing necessary components from `forest` object
+  object$forest$child.nodeIDs[which.trees] <- NULL
+  object$forest$split.values[which.trees] <- NULL
+  object$forest$split.varIDs[which.trees] <- NULL
+  object$forest$terminal.class.counts[which.trees] <- NULL  # for prob forests
+  object$forest$chf[which.trees] <- NULL  # for survival forests
+
+  # Update `num.trees` components so `predict.ranger()` works
+  object$forest$num.trees <- object$num.trees <- 
+    length(object$forest$child.nodeIDs)
+
+  # Coerce other components to `NA` as needed
+  if (!is.null(object$prediction.error)) {
+    object$prediction.error <- NA
+  } 
+  if (!is.null(object$predictions)) {  # classification and regression
+    object$predictions[] <- NA
+  }
+  if (!is.null(object$r.squared)) {  # regression
+    object$r.squared <- NA
+  }
+  if (!is.null(object$chf)) {  # survival forests
+    object$chf[] <- NA
+  }
+  if (!is.null(object$survival)) {  # survival forests
+    object$survival[] <- NA
+  }
+  if (object$importance.mode != "none") {  # variable importance
+    object$importance.mode <- NA
+    object$variable.importance[] <- NA
+  }
+
+  # Return "deforested" forest
+  class(object) <- c("deforest.ranger", class(object))
+  object
+
+}
+
+
+#' Print deforested ranger summary
+#' 
+#' Print basic information about a deforested \code{\link{ranger}} object.
+#' 
+#' @param x A \code{\link{deforest}} object (i.e., an object that inherits from
+#' class \code{"deforest.ranger"}).
+#' 
+#' @param ... Further arguments passed to or from other methods.
+#' 
+#' @note Many of the components of a typical \code{\link{ranger}} object are not 
+#' available after deforestation and are instead replaced with \code{NA} (e.g., 
+#' out-of-bag (OOB) predictions, variable importance scores (if requested), and 
+#' OOB-based error metrics).
+#' 
+#' @seealso \code{\link{deforest}}.
+#' 
+#' @author Brandon M. Greenwell
+#' 
+#' @export
+print.deforest.ranger <- function (x, ...) {
+  cat("Ranger (deforested) result\n\n")
+  cat("Note that many of the components of a typical \"ranger\" object are",
+      "not available after deforestation and are instead replaced with `NA`",
+      "(e.g., out-of-bag (OOB) predictions, variable importance scores (if",
+      "requested), and OOB-based error metrics)", 
+      "\n\n")
+  cat("Type:                            ", x$treetype, "\n")
+  cat("Number of trees:                 ", x$num.trees, "\n")
+  cat("Sample size:                     ", x$num.samples, "\n")
+  cat("Number of independent variables: ", x$num.independent.variables, "\n")
+  cat("Mtry:                            ", x$mtry, "\n")
+  cat("Target node size:                ", x$min.node.size, "\n")
+  cat("Variable importance mode:        ", x$importance.mode, "\n")
+  cat("Splitrule:                       ", x$splitrule, "\n")
+  if (x$treetype == "Survival") {
+    cat("Number of unique death times:    ", length(x$unique.death.times), "\n")
+  }
+  if (!is.null(x$splitrule) && x$splitrule == "extratrees" && 
+      !is.null(x$num.random.splits)) {
+    cat("Number of random splits:         ", x$num.random.splits, "\n")
+  }
+  if (x$treetype == "Classification") {
+    cat("OOB prediction error:            ", x$prediction.error, "\n")
+  }
+  else if (x$treetype == "Regression") {
+    cat("OOB prediction error (MSE):      ", x$prediction.error, "\n")
+  }
+  else if (x$treetype == "Survival") {
+    cat("OOB prediction error (1-C):      ", x$prediction.error, "\n")
+  }
+  else if (x$treetype == "Probability estimation") {
+    cat("OOB prediction error (Brier s.): ", x$prediction.error, "\n")
+  }
+  else {
+    cat("OOB prediction error:            ", x$prediction.error, "\n")
+  }
+  if (x$treetype == "Regression") {
+    cat("R squared (OOB):                 ", x$r.squared, "\n")
+  }
+}

tests/testthat/test_deforest.R

@@ -0,0 +1,43 @@
+library(ranger)
+library(survival)
+context("ranger_deforest")
+
+
+test_that("deforest works as expected for probability estimation", {
+  rfo <- ranger(Species ~ ., data = iris, num.trees = 10, probability = TRUE)
+  dfo <- deforest(rfo, which.trees = c(1, 3, 5), warn = FALSE)
+  pred.rfo <- predict(rfo, data = iris, predict.all = TRUE)$predictions
+  pred.dfo <- predict(dfo, data = iris, predict.all = TRUE)$predictions
+  expect_identical(pred.rfo[, , -c(1, 3, 5)], pred.dfo)
+})
+
+test_that("deforest works as expected for classification", {
+  rfo <- ranger(Species ~ ., data = iris, num.trees = 10)
+  dfo <- deforest(rfo, which.trees = c(1, 3, 5), warn = FALSE)
+  pred.rfo <- predict(rfo, data = iris, predict.all = TRUE)$predictions
+  pred.dfo <- predict(dfo, data = iris, predict.all = TRUE)$predictions
+  expect_identical(pred.rfo[, -c(1, 3, 5)], pred.dfo)
+})
+
+test_that("deforest works as expected for regression", {
+  n <- 50
+  x <- runif(n, min = 0, max = 2*pi)
+  dat <- data.frame(x = x, y = sin(x) + rnorm(n, sd = 0.1))
+  rfo <- ranger(y ~ ., data = dat, num.trees = 10)
+  dfo <- deforest(rfo, which.trees = c(1, 3, 5), warn = FALSE)
+  pred.rfo <- predict(rfo, data = dat, predict.all = TRUE)$predictions
+  pred.dfo <- predict(dfo, data = dat, predict.all = TRUE)$predictions
+  expect_identical(pred.rfo[, -c(1, 3, 5)], pred.dfo)
+})
+
+test_that("deforest works as expected for censored outcomes", {
+  dat <- data.frame(time = runif(100, 1, 10), status = rbinom(100, 1, .5), 
+                    x = rbinom(100, 1, .5))
+  rfo <- ranger(Surv(time, status) ~ x, data = dat, num.trees = 10, 
+                splitrule = "logrank")
+  dfo <- deforest(rfo, which.trees = c(1, 3, 5), warn = FALSE)
+  pred.rfo <- predict(rfo, data = dat, predict.all = TRUE)
+  pred.dfo <- predict(dfo, data = dat, predict.all = TRUE)
+  expect_identical(pred.rfo$chf[, , -c(1, 3, 5)], pred.dfo$chf)
+  expect_identical(pred.rfo$survival[, , -c(1, 3, 5)], pred.dfo$survival)
+})