MLearn-new: revised MLearn interface for machine learning

In lgatto/MLInterfaces: Uniform interfaces to R machine learning procedures for data in Bioconductor containers

Description

revised MLearn interface for machine learning, emphasizing a schematic description of external learning functions like knn, lda, nnet, etc.

Usage

MLearn( formula, data, .method, trainInd, ... )
makeLearnerSchema(packname, mlfunname, converter, predicter)

Arguments

formula	standard model formula
data	data.frame or ExpressionSet instance
.method	instance of learnerSchema
trainInd	obligatory numeric vector of indices of data to be used for training; all other data are used for testing, or instance of the xvalSpec class
...	additional named arguments passed to external learning function
packname	character – name of package harboring a learner function
mlfunname	character – name of function to use
converter	function – with parameters (obj, data, trainInd) that tells how to convert the material in obj [produced by [packname::mlfunname] ] into a classifierOutput instance.
predicter	function – with parameters (obj, newdata, ...) that tells how to use the material in obj to predict newdata.

Details

The purpose of the MLearn methods is to provide a uniform calling sequence to diverse machine learning algorithms. In R package, machine learning functions can have parameters (x, y, ...) or (formula, data, ...) or some other sequence, and these functions can return lists or vectors or other sorts of things. With MLearn, we always have calling sequence MLearn(formula, data, .method, trainInd, ...), and data can be a data.frame or ExpressionSet. MLearn will always return an S4 instance of classifierObject or clusteringObject.

At this time (1.13.x), NA values in predictors trigger an error.

To obtain documentation on the older (pre bioc 2.1) version of the MLearn method, please use help(MLearn-OLD).

randomForestI

randomForest. Note, that to obtain the default performance of randomForestB, you need to set mtry and sampsize parameters to sqrt(number of features) and table([training set response factor]) respectively, as these were not taken to be the function's defaults. Note you can use xvalSpec("NOTEST") as trainInd, to use all the samples; the RObject() result will print the misclassification matrix estimate along with OOB error rate estimate.

knnI(k=1,l=0)

knn; special support bridge required, defined in MLint

knn.cvI(k=1,l=0)

knn.cv; special support bridge required, defined in MLint. This option uses the embedded leave-one-out cross-validation of knn.cv, and thereby achieves high performance. You can have more general cross-validation using knnI with an xvalSpec, but it will be slower. When using this learner schema, you should use the numerical trainInd setting with 1:N where N is the number of samples.

dldaI

diagDA; special support bridge required, defined in MLint

nnetI

nnet

rpartI

rpart

ldaI

lda

svmI

svm

qdaI

qda

logisticI(threshold)

glm – with binomial family, expecting a dichotomous factor as response variable, not bulletproofed against other responses yet. If response probability estimate exceeds threshold, predict 1, else 0

adaI

ada

BgbmI

gbm, forcing the Bernoulli loss function.

blackboostI

blackboost – you MUST supply a family parameter relevant for mboost package procedures

lvqI

lvqtest after building codebook with lvqinit and updating with olvq1. You will need to write your own detailed schema if you want to tweak tuning parameters.

naiveBayesI

naiveBayes

baggingI

bagging

sldaI

slda

rdaI

rda – you must supply the alpha and delta parameters to use this. Typically cross-validation is used to select these. See rdacvI below.

rdacvI

rda.cv. This interface is complicated. The typical use includes cross-validation internal to the rda.cv function. That process searches a tuning parameter space and delivers an ordering on parameters. The interface selects the parameters by looking at all parameter configurations achieving the smallest min+1SE cv.error estimate, and taking the one among them that employed the -most- features (agnosticism). A final run of rda is then conducted with the tuning parameters set at that 'optimal' choice. The bridge code can be modified to facilitate alternative choices of the parameters in use. plotXvalRDA is an interface to the plot method for objects of class rdacv defined in package rda. You can use xvalSpec("NOTEST") with this procedure to use all the samples to build the discriminator.

ksvmI

ksvm

hclustI(distMethod, agglomMethod)

hclust – you must explicitly specify distance and agglomeration procedure.

kmeansI(centers, algorithm)

kmeans – you must explicitly specify centers and algorithm name.

If the parallel package is attached, cross-validation will be distributed to cores using mclapply.

Value

Instances of classifierOutput or clusteringOutput

Author(s)

Vince Carey <[e-mail]>

See Also

Try example(hclustWidget, ask=FALSE) for an interactive approach to cluster analysis tuning.

Examples

library("MASS")
data(crabs)
set.seed(1234)
kp = sample(1:200, size=120)
rf1 = MLearn(sp~CW+RW, data=crabs, randomForestI, kp, ntree=600 )
rf1
nn1 = MLearn(sp~CW+RW, data=crabs, nnetI, kp, size=3, decay=.01,
    trace=FALSE )
nn1
RObject(nn1)
knn1 = MLearn(sp~CW+RW, data=crabs, knnI(k=3,l=2), kp)
knn1
names(RObject(knn1))
dlda1 = MLearn(sp~CW+RW, data=crabs, dldaI, kp )
dlda1
names(RObject(dlda1))
lda1 = MLearn(sp~CW+RW, data=crabs, ldaI, kp )
lda1
names(RObject(lda1))
slda1 = MLearn(sp~CW+RW, data=crabs, sldaI, kp )
slda1
names(RObject(slda1))
svm1 = MLearn(sp~CW+RW, data=crabs, svmI, kp )
svm1
names(RObject(svm1))
ldapp1 = MLearn(sp~CW+RW, data=crabs, ldaI.predParms(method="debiased"), kp )
ldapp1
names(RObject(ldapp1))
qda1 = MLearn(sp~CW+RW, data=crabs, qdaI, kp )
qda1
names(RObject(qda1))
logi = MLearn(sp~CW+RW, data=crabs, glmI.logistic(threshold=0.5), kp, family=binomial ) # need family
logi
names(RObject(logi))
rp2 = MLearn(sp~CW+RW, data=crabs, rpartI, kp)
rp2
## recode data for RAB
#nsp = ifelse(crabs$sp=="O", -1, 1)
#nsp = factor(nsp)
#ncrabs = cbind(nsp,crabs)
#rab1 = MLearn(nsp~CW+RW, data=ncrabs, RABI, kp, maxiter=10)
#rab1
#
# new approach to adaboost
#
ada1 = MLearn(sp ~ CW+RW, data = crabs, .method = adaI, 
    trainInd = kp, type = "discrete", iter = 200)
ada1
confuMat(ada1)
#
lvq.1 = MLearn(sp~CW+RW, data=crabs, lvqI, kp )
lvq.1
nb.1 = MLearn(sp~CW+RW, data=crabs, naiveBayesI, kp )
confuMat(nb.1)
bb.1 = MLearn(sp~CW+RW, data=crabs, baggingI, kp )
confuMat(bb.1)
#
# new mboost interface -- you MUST supply family for nonGaussian response
#
require(party)  # trafo ... killing cmd check
blb.1 = MLearn(sp~CW+RW+FL, data=crabs, blackboostI, kp, family=mboost::Binomial() )
confuMat(blb.1)
#
# ExpressionSet illustration
# 
data(sample.ExpressionSet)
#  needed to increase training set size to avoid a new randomForest condition
# on empty class
set.seed(1234)
X = MLearn(type~., sample.ExpressionSet[100:250,], randomForestI, 1:19, importance=TRUE )
library(randomForest)
library(hgu95av2.db)
opar = par(no.readonly=TRUE)
par(las=2)
plot(getVarImp(X), n=10, plat="hgu95av2", toktype="SYMBOL")
par(opar)
#
# demonstrate cross validation
#
nn1cv = MLearn(sp~CW+RW, data=crabs[c(1:20,101:120),], 
   nnetI, xvalSpec("LOO"), size=3, decay=.01, trace=FALSE )
confuMat(nn1cv)
nn2cv = MLearn(sp~CW+RW, data=crabs[c(1:20,101:120),], nnetI, 
   xvalSpec("LOG",5, balKfold.xvspec(5)), size=3, decay=.01,
   trace=FALSE )
confuMat(nn2cv)
nn3cv = MLearn(sp~CW+RW+CL+BD+FL, data=crabs[c(1:20,101:120),], nnetI, 
   xvalSpec("LOG",5, balKfold.xvspec(5), fsFun=fs.absT(2)), size=3, decay=.01,
   trace=FALSE )
confuMat(nn3cv)
nn4cv = MLearn(sp~.-index-sex, data=crabs[c(1:20,101:120),], nnetI, 
   xvalSpec("LOG",5, balKfold.xvspec(5), fsFun=fs.absT(2)), size=3, decay=.01,
   trace=FALSE )
confuMat(nn4cv)
#
# try with expression data
#
library(golubEsets)
data(Golub_Train)
litg = Golub_Train[ 100:150, ]
g1 = MLearn(ALL.AML~. , litg, nnetI, 
   xvalSpec("LOG",5, balKfold.xvspec(5), 
   fsFun=fs.probT(.75)), size=3, decay=.01, trace=FALSE )
confuMat(g1)
#
# illustrate rda.cv interface from package rda (requiring local bridge)
#
library(ALL)
data(ALL)
#
# restrict to BCR/ABL or NEG
#
bio <- which( ALL$mol.biol %in% c("BCR/ABL", "NEG"))
#
# restrict to B-cell
#
isb <- grep("^B", as.character(ALL$BT))
kp <- intersect(bio,isb)
all2 <- ALL[,kp]
mads = apply(exprs(all2),1,mad)
kp = which(mads>1)  # get around 250 genes
vall2 = all2[kp, ]
vall2$mol.biol = factor(vall2$mol.biol) # drop unused levels

r1 = MLearn(mol.biol~., vall2, rdacvI, 1:40)
confuMat(r1)
RObject(r1)
plotXvalRDA(r1)  # special interface to plots of parameter space

# illustrate clustering support

cl1 = MLearn(~CW+RW+CL+FL+BD, data=crabs, hclustI(distFun=dist, cutParm=list(k=4)))
plot(cl1)

cl1a = MLearn(~CW+RW+CL+FL+BD, data=crabs, hclustI(distFun=dist, cutParm=list(k=4)), 
   method="complete")
plot(cl1a)

cl2 = MLearn(~CW+RW+CL+FL+BD, data=crabs, kmeansI, centers=5, algorithm="Hartigan-Wong")
plot(cl2, crabs[,-c(1:3)])

c3 = MLearn(~CL+CW+RW, crabs, pamI(dist), k=5)
c3
plot(c3, data=crabs[,c("CL", "CW", "RW")])


#  new interfaces to PLS thanks to Laurent Gatto

set.seed(1234)
kp = sample(1:200, size=120)

plsda.1 = MLearn(sp~CW+RW, data=crabs, plsdaI, kp, probMethod="Bayes")
plsda.1
confuMat(plsda.1)
confuMat(plsda.1,t=.65) ## requires at least 0.65 post error prob to assign species

plsda.2 = MLearn(type~., data=sample.ExpressionSet[100:250,], plsdaI, 1:16)
plsda.2
confuMat(plsda.2)
confuMat(plsda.2,t=.65) ## requires at least 0.65 post error prob to assign outcome

## examples for predict
clout <- MLearn(type~., sample.ExpressionSet[100:250,], svmI , 1:16)
predict(clout, sample.ExpressionSet[100:250,17:26])

lgatto/MLInterfaces documentation built on May 21, 2019, 5:12 a.m.