train: Fit Predictive Models over Different Tuning Parameters

Description

This function sets up a grid of tuning parameters for a number of classification and regression routines, fits each model and calculates a resampling based performance measure.

Usage

train(x, ...)
## S3 method for class 'default':
train(x, y, 
      method = "rf",  
      preProcess = NULL,
      ..., 
      weights = NULL,
      metric = ifelse(is.factor(y), "Accuracy", "RMSE"),   
      maximize = ifelse(metric == "RMSE", FALSE, TRUE),
      trControl = trainControl(), 
      tuneGrid = NULL, 
      tuneLength = 3)
## S3 method for class 'formula':
train(form, data, ..., weights, subset, na.action, contrasts = NULL)

Arguments

Value

A list is returned of class train containing:
methodthe chosen model.
modelTypean identifier of the model type.
resultsa data frame the training error rate and values of the tuning parameters.
bestTunea data frame with the final parameters.
callthe (matched) function call with dots expanded
dotsa list containing any ... values passed to the original call
metrica string that specifies what summary metric will be used to select the optimal model.
controlthe list of control parameters.
preProcesseither NULL or an object of class preProcess
finalModelan fit object using the best parameters
trainingDataa data frame
resampleA data frame with columns for each performance metric. Each row corresponds to each resample. If leave-one-out cross-validation or out-of-bag estimation methods are requested, this will be NULL. The returnResamp argument of trainControl controls how much of the resampled results are saved.
perfNamesa character vector of performance metrics that are produced by the summary function
maximizea logical recycled from the function arguments.
yLimitsthe range of the training set outcomes.
timesa list of execution times: everything is for the entire call to train, final for the final model fit and, optionally, prediction for the time to predict new samples (see trainControl)

Details

train can be used to tune models by picking the complexity parameters that are associated with the optimal resampling statistics. For particular model, a grid of parameters (if any) is created and the model is trained on slightly different data for each candidate combination of tuning parameters. Across each data set, the performance of held-out samples is calculated and the mean and standard deviation is summarized for each combination. The combination with the optimal resampling statistic is chosen as the final model and the entire training set is used to fit a final model.

A variety of models are currently available. The table below enumerates the models and the values of the method argument, as well as the complexity parameters used by train.

lccc{ Model method Value Package Tuning Parameter(s) Generalized linear model glm stats none glmStepAIC MASS none Generalized additive model gam mgcv select, method gamLoess gam span, degree gamSpline gam df Recursive partitioning rpart rpart cp rpart2 rpart maxdepth ctree party mincriterion ctree2 party maxdepth Boosted trees gbm gbm interaction depth, n.trees, shrinkage blackboost mboost maxdepth, mstop ada ada maxdepth, iter, nu bstTree bst maxdepth, mstop, nu Boosted regression models glmboost mboost mstop gamboost mboost mstop logitBoost caTools nIter bstLs bst mstop, nu bstSm bst mstop, nu Random forests rf randomForest mtry parRF randomForest, foreach mtry cforest party mtry Boruta Boruta mtry Bagging treebag ipred None bag caret vars logicBag logicFS ntrees, nleaves Other Trees nodeHarvest nodeHarvest maxinter, node partDSA partDSA cut.off.growth, MPD Logic Regression logreg LogicReg ntrees, treesize Elastic net (glm) glmnet glmnet alpha, lambda Neural networks nnet nnet decay, size neuralnet neuralnet layer1, layer2, layer3 pcaNNet caret decay, size avNNet caret decay, size, bag Projection pursuit regression ppr stats nterms Principal component regression pcr pls ncomp Independent component regression icr caret n.comp Partial least squares pls pls, caret ncomp simpls pls, caret ncomp widekernelpls pls, caret ncomp Sparse partial least squares spls spls, caret K, eta, kappa Support vector machines svmLinear kernlab C svmRadial kernlab sigma, C svmRadialCost kernlab C svmPoly kernlab scale, degree, C Relevance vector machines rvmLinear kernlab none rvmRadial kernlab sigma rvmPoly kernlab scale, degree Least squares support vector machines lssvmRadial kernlab sigma Gaussian processes guassprLinearl kernlab none guassprRadial kernlab sigma guassprPoly kernlab scale, degree Linear least squares lm stats None lmStepAIC MASS None leapForward leaps nvmax leapBackward leaps nvmax leapSeq leaps nvmax Robust linear regression rlm MASS None Multivariate adaptive regression splines earth earth degree, nprune gcvEarth earth degree Bagged MARS bagEarth caret, earth degree, nprune Rule Based Regression M5Rules RWeka pruned, smoothed M5 RWeka pruned, smoothed, rules cubist Cubist committees, neighbors Penalized linear models penalized penalized lambda1, lambda2 ridge elasticnet lambda enet elasticnet lambda, fraction lars lars fraction lars2 lars steps enet elasticnet fraction foba foba lambda, k Supervised principal components superpc superpc n.components, threshold Quantile regression forests qrf quantregForest mtry Quantile regression neural networks qrnn qrnn n.hidden, penalty, bag Linear discriminant analysis lda MASS None Linda rrcov None Quadratic discriminant analysis qda MASS None QdaCov rrcov None Stabilized linear discriminant analysis slda ipred None Heteroscedastic discriminant analysis hda hda newdim, lambda, gamma Stepwise discriminant analysis stepLDA klaR maxvar, direction stepQDA klaR maxvar, direction Stepwise diagonal discriminant analysis sddaLDA SDDA None sddaQDA SDDA None Shrinkage discriminant analysis sda sda diagonal Sparse linear discriminant analysis sparseLDA sparseLDA NumVars, lambda Regularized discriminant analysis rda klaR lambda, gamma Mixture discriminant analysis mda mda subclasses Sparse mixture discriminant analysis smda sparseLDA NumVars, R, lambda Penalized discriminant analysis pda mda lambda pda2 mda df Stabilised linear discriminant analysis slda ipred None High dimensional discriminant analysis hdda HDclassif model, threshold Flexible discriminant analysis (MARS) fda mda, earth degree, nprune Robust Regularized Linear Discriminant Analysis rrlda rrlda lambda, alpha Bagged FDA bagFDA caret, earth degree, nprune Logistic/multinomial regression multinom nnet decay Penalized logistic regression plr stepPlr lambda, cp Rule--based classification J48 RWeka C OneR RWeka None PART RWeka threshold, pruned JRip RWeka NumOpt Logic Forests logforest LogicForest None Bayesian multinomial probit model vbmpRadial vbmp estimateTheta k nearest neighbors knn3 caret k Nearest shrunken centroids pam pamr threshold scrda rda alpha, delta Naive Bayes nb klaR usekernel, fL Generalized partial least squares gpls gpls K.prov Learned vector quantization lvq class size, k ROC Curves rocc rocc xgenes }

By default, the function createGrid is used to define the candidate values of the tuning parameters. The user can also specify their own. To do this, a data fame is created with columns for each tuning parameter in the model. The column names must be the same as those listed in the table above with a leading dot. For example, ncomp would have the column heading .ncomp. This data frame can then be passed to createGrid.

In some cases, models may require control arguments. These can be passed via the three dots argument. Note that some models can specify tuning parameters in the control objects. If specified, these values will be superseded by those given in the createGrid argument.

The vignette entitled "caret Manual -- Model Building" has more details and examples related to this function.

train can be used with "explicit parallelism", where different resamples (e.g. cross-validation group) and models can be split up and run on multiple machines or processors. By default, train 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 foreach package. To run the resamples in parallel, the code for train does not change; prior to the call to train, a parallel backend is registered with foreach (see the examples below).

References

Kuhn (2008), ``Building Predictive Models in R Using the caret'' (http://www.jstatsoft.org/v28/i05/)

Examples

Run this code

#######################################
## Classification Example

data(iris)
TrainData <- iris[,1:4]
TrainClasses <- iris[,5]

knnFit1 <- train(TrainData, TrainClasses,
                 method = "knn",
                 preProcess = c("center", "scale"),
                 tuneLength = 10,
                 trControl = trainControl(method = "cv"))

knnFit2 <- train(TrainData, TrainClasses,
                 method = "knn",
                 preProcess = c("center", "scale"),
                 tuneLength = 10, 
                 trControl = trainControl(method = "boot"))


library(MASS)
nnetFit <- train(TrainData, TrainClasses,
                 method = "nnet",
                 preProcess = "range", 
                 tuneLength = 2,
                 trace = FALSE,
                 maxit = 100)

#######################################
## Regression Example

library(mlbench)
data(BostonHousing)

lmFit <- train(medv ~ . + rm:lstat,
               data = BostonHousing, 
               "lm")

library(rpart)
rpartFit <- train(medv ~ .,
                  data = BostonHousing,
                  "rpart",
                  tuneLength = 9)

#######################################
## Example with a custom metric

madSummary <- function (data,
                        lev = NULL,
                        model = NULL) 
{
  out <- mad(data$obs - data$pred, 
             na.rm = TRUE)  
  names(out) <- "MAD"
  out
}

robustControl <- trainControl(summaryFunction = madSummary)
marsGrid <- expand.grid(.degree = 1,
                        .nprune = (1:10) * 2)

earthFit <- train(medv ~ .,
                  data = BostonHousing, 
                  "earth",
                  tuneGrid = marsGrid,
                  metric = "MAD",
                  maximize = FALSE,
                  trControl = robustControl)

#######################################
## Parallel Processing Example via multicore package

library(doMC)
registerDoMC(2)

## The code for train() does not change:
set.seed(1)
usingMC <-  train(medv ~ .,
                  data = BostonHousing, 
                  "glmboost",
                  trControl = mcControl)

## or use:
## library(doMPI) or 
## library(doSMP) and so on

Run the code above in your browser using DataLab