%\VignetteIndexEntry{HOW TO use MCRestimate}
%\VignetteDepends{randomForest, e1071, xtable, RColorBrewer}
%\VignetteKeywords{Basic compendium - molecular markers - classification}
%\VignettePackage{MCRestimate}
\documentclass[12pt]{article}
\usepackage{times}
\usepackage{hyperref}
\usepackage[authoryear,round]{natbib}
\usepackage{times}
\usepackage{comment}
\textwidth=6.2in
\textheight=8.5in
%\parskip=.3cm
\oddsidemargin=.1in
\evensidemargin=.1in
\headheight=-.3in
\newcommand{\scscst}{\scriptscriptstyle}
\newcommand{\scst}{\scriptstyle}
\newcommand{\Rfunction}[1]{{\texttt{#1}}}
\newcommand{\Robject}[1]{{\texttt{#1}}}
\newcommand{\Rpackage}[1]{{\textit{#1}}}
\newcommand{\Rclass}[1]{{\textit{#1}}}
\newcommand{\Rmethod}[1]{{\textit{#1}}}
\newcommand{\Rfunarg}[1]{{\textit{#1}}}
\bibliographystyle{plainnat}
\begin{document}
\title{HOW TO use MCRestimate}
\author{Markus Ruschhaupt}
\maketitle
\section*{Estimation the misclassification error}
Every classification task starts with data. Here we choose the well
known ALL/AML data set from T.Golub~\cite{Golub.1999} availabel in the package \Rpackage{golubEsets}. Furthermore, we specify
the name of the phenodata column that should be used for the classification.
<<arguments,results=hide>>=
library(MCRestimate)
library(randomForest)
library(golubEsets)
data(Golub_Train)
class.colum <- "ALL.AML"
@
<<the functions, echo=FALSE,results=hide>>=
savepdf =function(x,file,w=10,h=5){pdf(file=file,width=w,height=w);x;dev.off()}
options(width=50)
@
Cross-validation is an appropiate instrument to estimate the misclassification rate
of an algorithm that should be used for classification. Often a variable
selection or aggregation is applied to the data set before the main
classification porcedure is started. But these steps must also be part of the
cross-validation. In our example we want to perform a variable selection and
only take the genes with the highest variance across all samples. Because we don't know exactly how many
genes we want to have, we give two possible values (250 and 1000). The described methods is
implemented in the functions \Rfunction{g.red.highest.var}. Further preprocessing functions are part of the
package \Rpackage{MCRestimate} and also new functions can be implemented.
<<METHOD CHOISE>>=
Preprocessingfunctions <- c("varSel.highest.var")
list.of.poss.parameter <- list(var.numbers=c(250,1000))
@
To use \Rfunction{MCRestimate} with a classification procedure a wrapper for this
method must be available. The package \Rpackage{MCRestimate} includes wrapper for the following classification
functions.
\begin{description}
\item[RF.wrap] wrapper for random forest (based on the package \Rpackage{randomForest})
\item[PAM.wrap] wrapper for PAM (based on the package \Rpackage{pamr})
\item[PLR.wrap] wrapper for the penalised logistic regression (based on the package \Rpackage{MCRestimate})
\item[SVM.wrap] wrapper for support vector machines (based on the package \Rpackage{e1071})
\item[GPLS.wrap] wrapper for generalised partial least squares (based on the package \Rpackage{gpls})
\end{description}
It is easy to write a wrapper for a new classification method. Here we want to use
random forest for our classification task.
<<METHOD CHOISE>>=
class.function <- "RF.wrap"
@
{\bf Parameter to optimize:} Most classification and preprocessing methods have parameters that must be
choosen and/or optimized. In our example we choose two possible number of
clusters for our preprocessing method. In each cross-validation step \Rfunction{MCRestimate} will
choose the value that have the lowes misclassification error on the training
set. This will be estimated through a second (inner) cross-validation with the
function tune available in the package \Rpackage{e1071}.
{\bf Parameter for the plot of the results (see man page for details):} We
want to have the sample names as labels in the resulting plot. {\textit Samples}
is the name of the phenoData column the sample names are stored in.
<<PLOT PARAMETER>>=
plot.label <- "Samples"
@
{\bf The Cross-validation parameter (see man page for details):} For time
reasons here we choose very small values. Normally the values for \Rfunarg{cross.outer} and
\Rfunarg{cross.inner} should be at least 5.
<<ARGUMENTS FOR CROSS-VALIDSATION>>=
cross.outer <- 2
cross.repeat <- 3
cross.inner <- 2
@
Now we have specified all parameter and can run the function
<<RF.make,eval=TRUE,results=hide>>=
RF.estimate <- MCRestimate(Golub_Train,
class.colum,
classification.fun="RF.wrap",
thePreprocessingMethods=Preprocessingfunctions,
poss.parameters=list.of.poss.parameter,
cross.outer=cross.outer,
cross.inner=cross.inner,
cross.repeat=cross.repeat,
plot.label=plot.label)
@
The result is an element of class \Rclass{MCRestimate}
<<rf.show>>=
class(RF.estimate)
@
For each group we see how many samples have been misclassified most of the time.
We can also visualize the result. The plot shows for each sample the number of
times it has been classified correctly.
<<RF,eval=FALSE,results=hide>>=
plot(RF.estimate,rownames.from.object=TRUE, main="Random Forest")
@
<<rf.save,echo=FALSE,results=hide>>=
savepdf(plot(RF.estimate,rownames.from.object=TRUE, main="Random Forest"),"RF.pdf")
@
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.7\textwidth]{RF}
\end{center}
\end{figure}
\section*{New data}
We have estimated the misclassification rate of our complete algorithm. Because this
seems to be a quit good result, we want to use this algorithm to classify new
data. The function \Rfunction{ClassifierBuild} is used to build a classifier that can be
used for new data.
<<>>=
RF.classifier <- ClassifierBuild (Golub_Train,
class.colum,
classification.fun="RF.wrap",
thePreprocessingMethods=Preprocessingfunctions,
poss.parameters=list.of.poss.parameter,
cross.inner=cross.inner)
@
The result of the function is a list with various arguments.
<<>>=
names(RF.classifier)
@
The most important elements of this list are \Rfunarg{classifier.for.matrix}
and \Rfunarg{classifier.for.exprSet}. These can now be used to classify
new data. The first one can be used if the new data is given by a matrix and
the second will be used if the new data is given by a \Rclass{exprSet}. Here we
want to classify the data from the \Rclass{exprSet} \Robject{golubTest}, that is
also part of the package \Rpackage{golubEsets}.
<<test>>=
data(Golub_Test)
RF.classifier$classifier.for.exprSet(Golub_Test)
@
This work was supported by NGFN (Nationales Genomforschungsnetz).
\begin{thebibliography}{}
\bibitem{Golub.1999} Golub TR, Slonim DK, Tamayo P, Huard C, Gaasenbeek M,
Mesirov JP, Coller H, Loh ML, Downing JR, Caligiuri MA, Bloomfield CD,
Lander ES.
\newblock Molecular classification of cancer: class discovery and class
prediction by gene expression monitoring
\newblock\textit{Science} 286(5439): 531-7 (1999).
\end{thebibliography}
\end{document}