fuse.default: Combining classifiers into a single ensemble model

In FuseBox: Combining Classifiers

Description

fuse provides a platform to take existing predictive models and combine there predictions into a single outcome.

Usage

fuse(mods, ...)

## Default S3 method:
fuse(mods,
             classes = NULL, 
             probs = TRUE, 
             predict = NULL, 
             weights = rep(1, length(mods)), 
             method = "vote", 
             methodArgs = NULL, 
             ...)

Arguments

mods	a named list of models
classes	a character string of possible classes
probs	a logical: will the model predict class probabilities (as opposed to the discrete class)
predict	an optional list the same length as mods that contains prediction functions for the models. By default, when probs = FALSE, samples are predicted using codepredict(model, newdata). When class probabilities are produced, the default syntax is predict(model, newdata, type = "prob"). The argument can be used for models that do not fit this convention or cases where the predictors do not use all the columns of newdata.
weights	a numeric vector the same length as mods of weights when averaging probabilities. These values will be normalized via weights/sum(weights).
method	usually a single method for combining the classifiers. Possible values are 'vote' (for majority vote), 'meanProb' (for weighted and unweighted averages of the class probabilities), 'prod' (the product of the class probabilities across models). Alternatively, a function with minimum arguments x and levels. See the Details section below.
methodArgs	an optional named list of arguments if a custom function is used with the method argument.
...	not currently used

Value

a list of class "fuse"

Author(s)

Max Kuhn

References

insert

See Also

predict.fuse, ~~~

Examples

## Not run: 
library(QSARdata)
data(Mutagen)

## Split the data three times into training and testing sets
library(caret)

set.seed(1)
inTrain <- createDataPartition(Mutagen_Outcome, p = 3/4, list = FALSE)

trainData <- Mutagen_Dragon[ inTrain,]
testData <- Mutagen_Dragon[-inTrain,]
trainClass <- Mutagen_Outcome[ inTrain]
testClass <- Mutagen_Outcome[-inTrain]

## There are some predictors with degenerate distirbutions, so 
## remvoe these

isNZV <- nearZeroVar(trainData)
trainData <- trainData[, -isNZV]
testData <- testData[, -isNZV]
calData <- calData[, -isNZV]

## Make a copy opf the data in a single data frame

training <- trainData
training$Class <- trainClass

## Fit a random forest model
library(randomForest)
rfModel <- randomForest(Class ~ ., data = training)

## Now an SVM model
library(kernlab)
svmModel <- ksvm(as.matrix(trainData), trainClass, C = 4, prob.model = TRUE)

## Create a list of the models and associated prediction functions:
models <- list(rf = rfModel,
               svm = svmModel)

pred <- list(function(x, dat, ...) predict(x, dat, type = "prob")[,1],
             function(x, dat, ...) predict(x, dat, type = "probabilities")[,1])

fusedMods <- fuse(list(rf = rfModel, svm = svmModel), 
                  probs = TRUE, 
                  predict = pred, 
                  method = "meanProb", 
                  classes = levels(testClass))

confusionMatrix(predict(fusedMods, testData), testClass)
confusionMatrix(predict(rfModel, testData), testClass)
confusionMatrix(predict(svmModel, testData), testClass)

## End(Not run)

FuseBox documentation built on May 2, 2019, 5:16 p.m.