Backwards Feature Selection
A simple backwards selection, a.k.a. recursive feature selection (RFE), algorithm
## S3 method for class 'default': rfe(x, y, sizes = 2^(2:4), metric = ifelse(is.factor(y), "Accuracy", "RMSE"), maximize = ifelse(metric == "RMSE", FALSE, TRUE), rfeControl = rfeControl(), ...)
rfeIter(x, y, testX, testY, sizes, rfeControl = rfeControl(), ...)
## S3 method for class 'rfe': predict(object, newdata, ...)
- a matrix or data frame of predictors for model training. This object must have unique column names.
- a vector of training set outcomes (either numeric or factor)
- a matrix or data frame of test set predictors. This must have the same column names as
- a vector of test set outcomes
- a numeric vector of integers corresponding to the number of features that should be retained
- a string that specifies what summary metric will be used to select the optimal model. By default, possible values are "RMSE" and "Rsquared" for regression and "Accuracy" and "Kappa" for classification. If custom performance metrics are used (via the
- a logical: should the metric be maximized or minimized?
- a list of options, including functions for fitting and prediction. See the package vignette or
- an object of class
- a matrix or data frame of new samples for prediction
- options to pass to the model fitting function (ignored in
This function implements backwards selection of predictors based on predictor importance ranking. The predictors are ranked and the less important ones are sequentially eliminated prior to modeling. The goal is to find a subset of predictors that can be used to produce an accurate model. The package vignette for feature selection has detailed descriptions of the algorithms.
rfe can be used with "explicit parallelism", where different resamples (e.g. cross-validation group) can be split up and run on multiple machines or processors. By default,
rfe will use a single processor on the host machine. As of version 4.99 of this package, the framework used for parallel processing uses the
rfe does not change; prior to the call to
rfe, a parallel backend is registered with
rfeIter is the basic algorithm while
rfe wraps these operations inside of resampling. To avoid selection bias, it is better to use the function
- A list with elements
finalVariables a list of size
length(sizes) + 1containing the column names of the ``surviving'' predictors at each stage of selection. The first element corresponds to all the predictors (i.e.
size = ncol(x))
pred a data frame with columns for the test set outcome, the predicted outcome and the subset size.
data(BloodBrain) x <- scale(bbbDescr[,-nearZeroVar(bbbDescr)]) x <- x[, -findCorrelation(cor(x), .8)] x <- as.data.frame(x) set.seed(1) lmProfile <- rfe(x, logBBB, sizes = c(2:25, 30, 35, 40, 45, 50, 55, 60, 65), rfeControl = rfeControl(functions = lmFuncs, number = 200)) set.seed(1) lmProfile2 <- rfe(x, logBBB, sizes = c(2:25, 30, 35, 40, 45, 50, 55, 60, 65), rfeControl = rfeControl(functions = lmFuncs, rerank = TRUE, number = 200)) xyplot(lmProfile$results$RMSE + lmProfile2$results$RMSE ~ lmProfile$results$Variables, type = c("g", "p", "l"), auto.key = TRUE) rfProfile <- rfe(x, logBBB, sizes = c(2, 5, 10, 20), rfeControl = rfeControl(functions = rfFuncs)) bagProfile <- rfe(x, logBBB, sizes = c(2, 5, 10, 20), rfeControl = rfeControl(functions = treebagFuncs)) set.seed(1) svmProfile <- rfe(x, logBBB, sizes = c(2, 5, 10, 20), rfeControl = rfeControl(functions = caretFuncs, number = 200), ## pass options to train() method = "svmRadial") ## classification with no resampling data(mdrr) mdrrDescr <- mdrrDescr[,-nearZeroVar(mdrrDescr)] mdrrDescr <- mdrrDescr[, -findCorrelation(cor(mdrrDescr), .8)] set.seed(1) inTrain <- createDataPartition(mdrrClass, p = .75, list = FALSE)[,1] train <- mdrrDescr[ inTrain, ] test <- mdrrDescr[-inTrain, ] trainClass <- mdrrClass[ inTrain] testClass <- mdrrClass[-inTrain] set.seed(2) ldaProfile <- rfe(train, trainClass, sizes = c(1:10, 15, 30), rfeControl = rfeControl(functions = ldaFuncs, method = "cv")) plot(ldaProfile, type = c("o", "g")) postResample(predict(ldaProfile, test), testClass) ####################################### ## Parallel Processing Example via multicore library(doMC) ## Note: if the underlying model also uses foreach, the ## number of cores specified above will double (along with ## the memory requirements) registerDoMC(cores = 2) set.seed(1) lmProfile <- rfe(x, logBBB, sizes = c(2:25, 30, 35, 40, 45, 50, 55, 60, 65), rfeControl = rfeControl(functions = lmFuncs, number = 200))