caret (version 4.51)

bag.default: A General Framework For Bagging

Description

bag provides a framework for bagging classification or regression models. The user can provide their own functions for model building, prediction and aggregation of predictions (see Details below).

Usage

bag(x, ...)
## S3 method for class 'default':
bag(x, y, B = 10, vars = NULL, bagControl = bagControl(), ...)
bagControl(fit = NULL, 
           predict = NULL, 
           aggregate = NULL)

## S3 method for class 'bag':
predict(object, newdata = NULL, ...)

Arguments

x

a matrix or data frame of predictors

y

a vector of outcomes

B

the number of bootstrap samples to train over.

bagControl

a list of options.

...

arguments to pass to the model function

fit

a function that has arguments x, y and ... and produces a model object that can later be used for prediction

predict

a function that generates predictions for each sub-model. The function should have arguments object and x. The output of the function can be any type of object (see the example below where posterior probabilities are generated)

aggregate

a function with arguments x and type. The function that takes the output of the predict function and reduces the bagged predictions to a single prediction per sample. the type argument can be used to swi

vars

an integer. If this argument is not NULL, a random sample of size vars is taken of the predictors in each bagging iteration. If NULL, all predictors are used.

object

an object of class bag.

newdata

a matrix or data frame of samples for prediction. Note that this argument must have a non-null value

Value

• bag produces an object of class bag with elements
• fits
  a list with two sub-objects: the fit object has the actual model fit for that bagged samples and the vars object is either NULL or a vector of integers corresponding to which predictors were sampled for that model
• control
  a mirror of the arguments passed into bagControl
• call
  the call
• B
  the number of bagging iterations
• dims
  the dimensions of the training set

Details

The function is basically a framework where users can plug in any model in to assess the effect of bagging.

One note: when vars is not NULL, the sub-setting occurs prior to the fit and predict functions are called. In this way, the user probably does not need to account for the change in predictors in their functions.

When using bag with train, classification models should use type = "prob" inside of the predict function so that predict.train(object, newdata, type = "prob") will work.

Examples

## A simple example of bagging conditional inference regression trees:
data(BloodBrain)

## Fit a model with the default values
ctreeFit <- function(x, y, ...)
  {
    library(party)
    data <- as.data.frame(x)
    data$y <- y
    ctree(y~., data = data)
  }

## Generate simple predictions of the outcome
ctreePred <- function(object, x)
  {
    predict(object, x)[,1]
  }

## Take the median of the bagged predictions
ctreeAg <- function(x, type = NULL)
  {
    ## x is a list of vectors, so we convert them to a matrix
    preds <- do.call("cbind", x)
    apply(preds, 1, median)
  }

treebag <- bag(bbbDescr, logBBB, B = 10,
               bagControl = bagControl(fit = ctreeFit,
                                       predict = ctreePred,
                                       aggregate = ctreeAg))




## An example of pooling posterior probabilities to generate class predictions
data(mdrr)

## remove some zero variance predictors and linear dependencies
mdrrDescr <- mdrrDescr[, -nearZeroVar(mdrrDescr)]
mdrrDescr <- mdrrDescr[, -findCorrelation(cor(mdrrDescr), .95)]


## The fit and predict functions are stright-forward:
ldaFit <- function(x, y, ...)
  {
    library(MASS)
    lda(x, y, ...)
  }

ldaPred <- function(object, x)
  {
    predict(object, x)$posterior
  }


## For the aggregation function, we take the median of the bagged
## posterior probabilities and pick the largest as the class
ldaAg <- function(x, type = "class")
  {
    ## The class probabilities come in as a list of matrices
    ## For each class, we can pool them then average over them

    pooled <- x[[1]] & NA
    classes <- colnames(pooled)
    for(i in 1:ncol(pooled))
      {
        tmp <- lapply(x, function(y, col) y[,col], col = i)
        tmp <- do.call("rbind", tmp)
        pooled[,i] <- apply(tmp, 2, median)
      }
    if(type == "class")
      {
        out <- factor(classes[apply(pooled, 1, which.max)],
                      levels = classes)
      } else out <- pooled
    out
  }


bagLDA <- bag(mdrrDescr, mdrrClass, 
              B = 10,
              vars = 10,
              bagControl = bagControl(fit = ldaFit,
                                      predict = ldaPred,
                                      aggregate = ldaAg))

basicLDA <- train(mdrrDescr, mdrrClass, "lda")

bagLDA2 <- train(mdrrDescr, mdrrClass, 
                 "bag", 
                 B = 10, 
                 bagControl(fit = ldaFit,
                            predict = ldaPred,
                            aggregate = ldaAg),
                 tuneGrid = data.frame(.vars = c((1:10)*10 , ncol(mdrrDescr))))