mlDNA (version 1.1)

classifier: Building machine learning-based classification models

Description

This function builds classification models with different machine learning algorithms including random forest (randomForest), support vector machine (svm), and neural network (nnet).

Usage

classifier( method = c("randomForest", "svm", "nnet" ),  featureMat, positiveSamples, negativeSamples,  tunecontrol = tune.control(sampling = "cross", cross = 5), ...)

Arguments

method

a character string specifying machine learning method used to construct prediction models. Currently the user-optional values are "randomForest", "svm" or "nnet".

featureMat

a numeric matrix recording feature values for each sample.

positiveSamples

a character vector recording positive samples used to train classification model.

negativeSamples

a character vector recording negative samples used to train classification model.

tunecontrol

control parameters for the tune function. The value is assigned by the function "tune.control" in e1071 package.

...

Further parameters passed to train classification model.

Value

An object of class tune (see tune function in e1071 package) is returned, including the components:

best.parameters

a 1 x k data frame, k number of parameters.

best.performance

best achieved performance.

best.model

the model trained on the complete training data using the best parameter combination.

Details

For the random forest algorithm, the important parameters are mtry (number of features randomly selected for bulding the decision tree. Default:sqrt(ncol(featureMat))) and ntree (number of trees to be built. Default:500). More information about the parameters related to random forest can be found in R package RandomForest.

For the svm algorithm, the default kernel function is the "radial"(radial basis function; RBF). Other optional kernels are the "linear", the "polynomial", and the "sigmoid" fuctions. The range of degree (for the kernel type of polynomial), gamma (for radial), coef0 (for polynomial and sigmoid) and cost (the cost parameters for all kernels) should be designated. Please refer the description of "svm" function in R package e1071 for more information about the parameters of svm.

For the neural network, the important parameters are size (number of units in the hidden layer), decay (parameter for weight decay. Default 0), trace (switch for tracing optimization. Default TRUE.). More information about the parameters of neural network is described in nnet and e1071 packages.

Examples

## Not run: 
# 
#    ##generate expression feature matrix
#    sampleVec1 <- c(1, 2, 3, 4, 5, 6)
#    sampleVec2 <- c(1, 2, 3, 4, 5, 6)
#    featureMat <- expFeatureMatrix( expMat1 = ControlExpMat, sampleVec1 = sampleVec1, 
#                                    expMat2 = SaltExpMat, sampleVec2 = sampleVec2, 
#                                    logTransformed = TRUE, base = 2,
#                                    features = c("zscore", "foldchange", "cv", "expression"))
# 
#    ##positive samples
#    positiveSamples <- as.character(sampleData$KnownSaltGenes)
#    ##unlabeled samples
#    unlabelSamples <- setdiff( rownames(featureMat), positiveSamples )
#    idx <- sample(length(unlabelSamples))
#    ##randomly selecting 1000 unlabeled samples as negative samples
#    negativeSamples <- unlabelSamples[idx[1:1000]]
# 
#    ##for random forest, and using five-fold cross validation for obtaining optimal parameters
#    cl <- classifier( method = "randomForest", 
#                      featureMat = featureMat, 
#                      positiveSamples = positiveSamples, 
#                      negativeSamples = negativeSamples,
#                      tunecontrol = tune.control(sampling = "cross", cross = 5),
#                      ntree = 100 ) #build 100 trees for the forest
# 
# 
#    ##svm and using five-fold cross validation for obtaining optimal parameters
#    cl <- classifier( method = "svm", featureMat = featureMat, 
#                      positiveSamples = positiveSamples, 
#                      negativeSamples = negativeSamples,
#                      tunecontrol = tune.control(sampling = "cross", cross = 5),
#                      kernel = "radial", 
#                      probability = TRUE,
#                      ranges = list(gamma = 2^(-2:2), 
#                       cost = 2^(-4:4)) ) #for radial kernel and the parameter space.
# 
#    ##neural network, using one split for training/validation set
#    cl <- classifier( method = "nnet", featureMat = featureMat, 
#                      positiveSamples = positiveSamples, 
#                      negativeSamples = negativeSamples,
#                      tunecontrol = tune.control(sampling = "fix"), 
#                      trace = TRUE, size = 10 ) #for nnet parameters.
#   
# 
# ## End(Not run)