| Title: | Nested Loop Cross Validation |
|---|---|
| Description: | Nested loop cross validation for classification purposes for misclassification error rate estimation. The package supports several methodologies for feature selection: random forest, Student t-test, limma, and provides an interface to the following classification methods in the 'MLInterfaces' package: linear, quadratic discriminant analyses, random forest, bagging, prediction analysis for microarray, generalized linear model, support vector machine (svm and ksvm). Visualizations to assess the quality of the classifier are included: plot of the ranks of the features, scores plot for a specific classification algorithm and number of features, misclassification rate for the different number of features and classification algorithms tested and ROC plot. For further details about the methodology, please check: Markus Ruschhaupt, Wolfgang Huber, Annemarie Poustka, and Ulrich Mansmann (2004) <doi:10.2202/1544-6115.1078>. |
| Authors: | Willem Talloen, Tobias Verbeke |
| Maintainer: | Laure Cougnaud <[email protected]> |
| License: | GPL-3 |
| Version: | 0.3.5 |
| Built: | 2024-10-26 06:43:55 UTC |
| Source: | CRAN |
function to compare the original matrix of correct classes to each component of the output object for a certain classifier
compareOrig(nlcvObj, techn)compareOrig(nlcvObj, techn)
nlcvObj |
return of the |
techn |
technique for which the comparison to correct classes should be made |
list with for each number of features selected, a matrix of logical values indicating whether the classifier results correspond (TRUE) or not (FALSE) to the original values to be classified
The observed and predicted classes are cross-tabulated for a given
classification technique used in the nested loop cross validation. The
predicted class that is used to construct the confusion matrix is the class
that was predicted most of the time () across all runs
of the nested loop.
## S3 method for class 'nlcv' confusionMatrix(x, tech, proportions = TRUE, ...)## S3 method for class 'nlcv' confusionMatrix(x, tech, proportions = TRUE, ...)
x |
object for which a confusionMatrix should be produced, e.g. one
produced by the |
tech |
string indicating the classification technique for which the confusion matrix should be returned |
proportions |
logical indicating whether the cells of the matrix should
contain proportions ( |
... |
Dots argument to pass additional parameters to the
|
confusionMatrix produces an object of class
confusionMatrix which directly inherits from the ftable class
(representing the confusion matrix)
Willem Talloen and Tobias Verbeke
This function takes in a factor with class labels of the total dataset,
draws a sample (balanced with respect to the different levels of the factor)
and returns a logical vector indicating whether the observation is in the
learning sample (TRUE) or not (FALSE).
inTrainingSample(y, propTraining = 2/3, classdist = c("balanced", "unbalanced"))inTrainingSample(y, propTraining = 2/3, classdist = c("balanced", "unbalanced"))
y |
factor with the class labels for the total data set |
propTraining |
proportion of the data that should be in a training set; the default value is 2/3. |
classdist |
distribution of classes; allows to indicate whether your distribution 'balanced' or 'unbalanced'. The sampling strategy for each run is adapted accordingly. |
logical vector indicating for each observation in y whether
the observation is in the learning sample (TRUE) or not
(FALSE)
Willem Talloen and Tobias Verbeke
### this example demonstrates the logic of sampling in case of unbalanced distribution of classes y <- factor(c(rep("A", 21), rep("B", 80))) nlcv:::inTrainingSample(y, 2/3, "unbalanced") table(y[nlcv:::inTrainingSample(y, 2/3, "unbalanced")]) # should be 14, 14 (for A, B resp.) table(y[!nlcv:::inTrainingSample(y, 2/3, "unbalanced")]) # should be 7, 66 (for A, B resp.)### this example demonstrates the logic of sampling in case of unbalanced distribution of classes y <- factor(c(rep("A", 21), rep("B", 80))) nlcv:::inTrainingSample(y, 2/3, "unbalanced") table(y[nlcv:::inTrainingSample(y, 2/3, "unbalanced")]) # should be 14, 14 (for A, B resp.) table(y[!nlcv:::inTrainingSample(y, 2/3, "unbalanced")]) # should be 7, 66 (for A, B resp.)
Wrapper around limma for the comparison of two groups
limmaTwoGroups(object, group)limmaTwoGroups(object, group)
object |
object of class ExpressionSet |
group |
string indicating the variable defining the two groups to be compared |
Basically, the wrapper combines the lmFit, eBayes and
topTable steps
topTable output for the second (i.e. slope) coefficient of
the linear model.
Tobias Verbeke
Smyth, G. K. (2004). Linear models and empirical Bayes methods for assessing differential expression in microarray experiments. Statistical Applications in Genetics and Molecular Biology, Vol. 3, No. 1, Article 3.
http://www.bepress.com/sagmb/vol3/iss1/art3
plots for each classification technique and a given number of features used the mean misclassification rate (mcr) and its standard error across all runs of the nested loop cross-validation.
mcrPlot(nlcvObj, plot = TRUE, optimalDots = TRUE, rescale = FALSE, layout = TRUE, ...)mcrPlot(nlcvObj, plot = TRUE, optimalDots = TRUE, rescale = FALSE, layout = TRUE, ...)
nlcvObj |
Object of class 'nlcv' as produced by the |
plot |
logical. If |
optimalDots |
Boolean indicating whether dots should be displayed on a panel below the graph to mark the optimal number of features for a given classification technique |
rescale |
if |
layout |
boolean indicating whether |
... |
Dots argument to pass additional graphical parameters (such as
|
An MCR plot is output to the device of choice. The dots represent the mean MCR across runs. The vertical lines below and above the dots represent the standard deviation of the MCR values across runs.
Below the plot coloured solid dots (one for each classification technique) indicate for which number of features a given technique reached its minimum MCR.
The function invisibly returns an object of class mcrPlot which is a
list with components:
meanMcrMatrixmatrix with for each number of features (rows) and classification technique (columns) the mean of the MCR values across all runs of the nlcv procedure.
sdMcrMatrixmatrix with for each number of features (rows) and classification technique (columns) the sd of the MCR values across all runs of the nlcv procedure.
The summary method for the mcrPlot object returns a matrix
with for each classification technique, the optimal number of features as
well as the associated mean MCR and standard deviation of the MCR values.
Willem Talloen and Tobias Verbeke
This function first proceeds to a feature selection and then applies five different classification algorithms.
nlcv(eset, classVar = "type", nRuns = 2, propTraining = 2/3, classdist = c("balanced", "unbalanced"), nFeatures = c(2, 3, 5, 7, 10, 15, 20, 25, 30, 35), fsMethod = c("randomForest", "t.test", "limma", "none"), classifMethods = c("dlda", "randomForest", "bagg", "pam", "svm"), fsPar = NULL, initialGenes = seq(length.out = nrow(eset)), geneID = "ID", storeTestScores = FALSE, verbose = FALSE, seed = 123)nlcv(eset, classVar = "type", nRuns = 2, propTraining = 2/3, classdist = c("balanced", "unbalanced"), nFeatures = c(2, 3, 5, 7, 10, 15, 20, 25, 30, 35), fsMethod = c("randomForest", "t.test", "limma", "none"), classifMethods = c("dlda", "randomForest", "bagg", "pam", "svm"), fsPar = NULL, initialGenes = seq(length.out = nrow(eset)), geneID = "ID", storeTestScores = FALSE, verbose = FALSE, seed = 123)
eset |
ExpressionSet object containing the genes to classify |
classVar |
String giving the name of the variable containing the
observed class labels, should be contained in the phenoData of |
nRuns |
Number of runs for the outer loop of the cross-validation |
propTraining |
Proportion of the observations to be assigned to the
training set. By default |
classdist |
distribution of classes; allows to indicate whether your distribution is 'balanced' or 'unbalanced'. The sampling strategy for each run is adapted accordingly. |
nFeatures |
Numeric vector with the number of features to be selected from the features kept by the feature selection method. For each number n specified in this vector the classification algorithms will be run using only the top n features. |
fsMethod |
Feature selection method; one of |
classifMethods |
character vector with the classification methods to be
used in the analysis; elements can be chosen among
|
fsPar |
List of further parameters to pass to the feature selection
method; currently the default for |
initialGenes |
Initial subset of genes in the ExpressionSet on which to apply the nested loop cross validation procedure. By default all genes are selected. |
geneID |
string representing the name of the gene ID variable in the fData of the expression set to use; this argument was added for people who use e.g. both Entrez IDs and Ensemble gene IDs |
storeTestScores |
should the test scores be stored in the |
verbose |
Should the output be verbose ( |
seed |
integer with seed, set at the start of the cross-validation. |
The result is an object of class 'nlcv'. It is a list with two
components, output and features.
De output component is a list of five components, one for each
classification algorithm used. Each of these components has as many
components as there are elements in the nFeatures vector. These
components contain both the error rates for each run (component
errorRate) and the predicted labels for each run (character matrix
labelsMat).
The features list is a list with as many components as there are
runs. For each run, a named vector is given with the variable importance
measure for each gene. For t test based feature selection, P-values are
used; for random forest based feature selection the variable importance
measure is given.
The variable importance measure used is the third column of the output
returned by the randomForest function.
Willem Talloen and Tobias Verbeke
This data set contains the nlcv results of selection of features with random forest on a randomly generated dataset.
nlcvRF_RnlcvRF_R
nlcv object
This data set contains the nlcv results of selection of features with random forest on a dataset with strong hetero signal.
nlcvRF_SHSnlcvRF_SHS
nlcv object
This data set contains the nlcv results of selection of features with random forest on a dataset with strong signal.
nlcvRF_SSnlcvRF_SS
nlcv object
This data set contains the nlcv results of selection of features with random forest on a weak signal dataset.
nlcvRF_WHSnlcvRF_WHS
nlcv object
This data set contains the nlcv results of selection of features with random forest on a weak hetero signal dataset.
nlcvRF_WSnlcvRF_WS
nlcv object
This data set contains the nlcv results of selection of features with t-test on a randomly generated dataset.
nlcvTT_RnlcvTT_R
nlcv object
This data set contains the nlcv results of selection of features with t-test on a dataset with strong hetero signal.
nlcvTT_SHSnlcvTT_SHS
nlcv object
This data set contains the nlcv results of selection of features with t-test on a dataset with strong signal.
nlcvTT_SSnlcvTT_SS
nlcv object
This data set contains the nlcv results of selection of features with t-test on a weak signal dataset.
nlcvTT_WHSnlcvTT_WHS
nlcv object
This data set contains the nlcv results of selection of features with t-test on a weak hetero signal dataset.
nlcvTT_WSnlcvTT_WS
nlcv object
This interface keeps track of the predictions on the training and test set, contrary to the ldaI interface that is made available in the MLInterfaces package.
nldaInldaI
An object of class learnerSchema of length 1.
nldaI is an object of class 'learnerSchema' and can be used as such in calls to MLearn (from MLInterfaces).
See Also ldaI
This object is an instance of the learnerSchema object and will be typically
used as the method argument of an MLearn call.
pamrIpamrI
An object of class learnerSchema of length 1.
Tobias Verbeke
set.seed(120) x <- matrix(rnorm(1000*20), ncol=20) y <- sample(c(1:4), size=20, replace=TRUE) alldf <- cbind.data.frame(t(x), y) # assure it is a factor (otherwise error message) alldf$y <- factor(alldf$y) library(MLInterfaces) (mlobj <- MLearn(y ~ ., data = alldf, .method = pamrI, trainInd = 1:15))set.seed(120) x <- matrix(rnorm(1000*20), ncol=20) y <- sample(c(1:4), size=20, replace=TRUE) alldf <- cbind.data.frame(t(x), y) # assure it is a factor (otherwise error message) alldf$y <- factor(alldf$y) library(MLInterfaces) (mlobj <- MLearn(y ~ ., data = alldf, .method = pamrI, trainInd = 1:15))
pamrML to classifierOutput
convert from pamrML to classifierOutput
pamrIconverter(obj, data, trainInd)pamrIconverter(obj, data, trainInd)
obj |
object as returned by pamrML i.e. of class |
data |
original data used as input for MLearn |
trainInd |
training indices used as input to MLearn |
object of class classifierOutput
The pamrML functions are wrappers around pamr.train and
pamr.predict that provide a more classical R modelling interface than
the original versions.
pamrML(formula, data, ...)pamrML(formula, data, ...)
formula |
model formula |
data |
data frame |
... |
argument for the |
The name of the response variable is kept as an attribute in the
pamrML object to allow for predict methods that can be easily used
for writing converter functions for use in the MLInterfaces
framework.
For pamrML an object of class pamrML which adds an
attribute to the original object returned by pamr.train (or
pamrTrain).
The print method lists the names of the different components of the
pamrML object.
The predict method returns a vector of predicted values
Tobias Verbeke
set.seed(120) x <- matrix(rnorm(1000*20), ncol=20) y <- sample(c(1:4), size=20, replace=TRUE) # for original pam mydata <- list(x=x, y=y) mytraindata <- list(x=x[,1:15],y=factor(y[1:15])) mytestdata <- list(x = x[,16:20], y = factor(y[16:20])) # for formula-based methods including pamrML alldf <- cbind.data.frame(t(mydata$x), y) traindf <- cbind.data.frame(t(mytraindata$x), y = mytraindata$y) testdf <- cbind.data.frame(t(mytestdata$x), y = mytestdata$y) ### create pamrML object pamrMLObj <- pamrML(y ~ ., traindf) pamrMLObj ### test predict method predict(object = pamrMLObj, newdata = testdf, threshold = 1) # threshold compulsoryset.seed(120) x <- matrix(rnorm(1000*20), ncol=20) y <- sample(c(1:4), size=20, replace=TRUE) # for original pam mydata <- list(x=x, y=y) mytraindata <- list(x=x[,1:15],y=factor(y[1:15])) mytestdata <- list(x = x[,16:20], y = factor(y[16:20])) # for formula-based methods including pamrML alldf <- cbind.data.frame(t(mydata$x), y) traindf <- cbind.data.frame(t(mytraindata$x), y = mytraindata$y) testdf <- cbind.data.frame(t(mytestdata$x), y = mytestdata$y) ### create pamrML object pamrMLObj <- pamrML(y ~ ., traindf) pamrMLObj ### test predict method predict(object = pamrMLObj, newdata = testdf, threshold = 1) # threshold compulsory
Function that provides a classical R modelling interface, using a
formula and data argument
pamrTrain(formula, data, ...)pamrTrain(formula, data, ...)
formula |
formula |
data |
data frame |
... |
further arguments to be passed to |
Object that is perfectly identical to the object returned by
pamr.train
Tobias Verbeke
set.seed(120) x <- matrix(rnorm(1000*20), ncol=20) y <- sample(c(1:4), size=20, replace=TRUE) alldf <- cbind.data.frame(t(x), y) pamrTrain(y ~ ., alldf)set.seed(120) x <- matrix(rnorm(1000*20), ncol=20) y <- sample(c(1:4), size=20, replace=TRUE) alldf <- cbind.data.frame(t(x), y) pamrTrain(y ~ ., alldf)
pamrML objectpredict pamrML object
## S3 method for class 'pamrML' predict(object, newdata, ...)## S3 method for class 'pamrML' predict(object, newdata, ...)
object |
|
newdata |
new data |
... |
additional parameters for the pamr.predict function |
output of the pamr.predict function
nlcvConfusionMatrix
print object nlcvConfusionMatrix
## S3 method for class 'nlcvConfusionMatrix' print(x, ...)## S3 method for class 'nlcvConfusionMatrix' print(x, ...)
x |
object of class |
... |
additional parameters for the |
no returned value, the object is printed in the output
pamrML objectprint pamrML object
## S3 method for class 'pamrML' print(x, ...)## S3 method for class 'pamrML' print(x, ...)
x |
object of class |
... |
additional parameters for the |
print function for summary.mcrPlot objectprint function for summary.mcrPlot object
## S3 method for class 'summary.mcrPlot' print(x, digits = 2, ...)## S3 method for class 'summary.mcrPlot' print(x, digits = 2, ...)
x |
Object of class 'summary.mcrPlot' as produced by the function of the same name |
digits |
number of digits to be passed to the default print method |
... |
additional parameters for the |
This plot offers an overview of the distribution of the ranks of the n best-ranked features. The order of the features is determined by the median rank of the feature across all nlcv runs.
rankDistributionPlot(nlcvObj, n = 5, ...)rankDistributionPlot(nlcvObj, n = 5, ...)
nlcvObj |
object of class |
n |
number of features for whicht the distribution should be displayed |
... |
additional arguments to the boxplot functions (such as
|
For each of the n features, a boxplot is displayed.
Willem Talloen and Tobias Verbeke
{ data(nlcvRF_SS) rankDistributionPlot(nlcvRF_SS, n = 9) }{ data(nlcvRF_SS) rankDistributionPlot(nlcvRF_SS, n = 9) }
Produce a ROC plot for a classification model belonging to a given technique and with a given number of features.
rocPlot(nlcvObj, tech, nfeat, main = NULL, globalAUCcol = "#FF9900", ...)rocPlot(nlcvObj, tech, nfeat, main = NULL, globalAUCcol = "#FF9900", ...)
nlcvObj |
object of class 'nlcv' as produced by the nlcv function |
tech |
technique; character of length one; one of 'dlda', 'lda', 'nlda', 'qda', 'glm', 'randomForest', 'bagg', 'pam', 'svm' or 'ksvm' |
nfeat |
number of features used in the classification model; numeric of length one |
main |
main title to be used for the ROC plot |
globalAUCcol |
color for the global AUC (defaults to '#FF9900') |
... |
further arguments for the plot call (such as sub e.g.) |
A ROC plot is drawn to the current device
Tobias Verbeke
Function to plot, for a given nested loop cross-validation object, a given classification technique and a given number of features used for the classification, the scores plot. This plot diplays the proportion of correctly-classified per sample across all runs of the nested loop cross-validation. The class membership of the samples is displayed using a colored strip (with legend below the plot).
scoresPlot(nlcvObj, tech, nfeat, plot = TRUE, barPlot = FALSE, layout = TRUE, main = NULL, sub = NULL, ...)scoresPlot(nlcvObj, tech, nfeat, plot = TRUE, barPlot = FALSE, layout = TRUE, main = NULL, sub = NULL, ...)
nlcvObj |
Object of class 'nlcv' as produced by the |
tech |
string denoting the classification technique used; one of 'dlda', 'bagg', 'pam', 'rf', or 'svm'. |
nfeat |
integer giving the number of features; this number should be
part of the initial set of number of features that was specified during the
nested loop cross-validation ( |
plot |
logical. If |
barPlot |
Should a barplot be drawn ( |
layout |
boolean indicating whether |
main |
Main title for the scores plot; if not supplied, 'Scores Plot' is used as a default |
sub |
Subtitle for the scores plot; if not supplied, the classification technique and the chosen number of features are displayed |
... |
Additional graphical parameters to pass to the plot function |
A scores plot is displayed (for the device specified).
The function invisibly returns a named vector containing (for each sample) the proportion of times the sample was correctly classified (for a given technique and a given number of features used).
Willem Talloen and Tobias Verbeke
summary function for mcrPlot objectsummary function for mcrPlot object
## S3 method for class 'mcrPlot' summary(object, ...)## S3 method for class 'mcrPlot' summary(object, ...)
object |
Object of class 'mcrPlot' as produced by the function of the same name |
... |
additional arguments, not used here |
Methods for topTable. topTable extracts the top n most important features for a given classification or regression procedure.
## S4 method for signature 'nlcv' topTable(fit, n = 5, method = "percentage")## S4 method for signature 'nlcv' topTable(fit, n = 5, method = "percentage")
fit |
object resulting from a classification or regression procedure |
n |
number of features that one wants to extract from a table that ranks all features according to their importance in the classification or regression model |
method |
method used to rank the features; one of |
The top n features are extracted across all runs of the nested loop cross-validation. After ranking on their frequency of selection, the top n are retained and returned.
a data frame of one column (percentage) with percentages
reflecting the frequency of selection of a feature in the top n across all
runs; the features are sorted on decreasing frequency.
nlcv
nlcv objects are produced by nlcv
Willem Talloen and Tobias Verbeke
data(nlcvRF_SS) topTable(nlcvRF_SS, n = 7, method = "medianrank")data(nlcvRF_SS) topTable(nlcvRF_SS, n = 7, method = "medianrank")
xtable method for confusionMatrix objects
## S3 method for class 'confusionMatrix' xtable(x, caption = NULL, label = NULL, align = NULL, digits = NULL, display = NULL, ...)## S3 method for class 'confusionMatrix' xtable(x, caption = NULL, label = NULL, align = NULL, digits = NULL, display = NULL, ...)
x |
object of class 'confusionMatrix' as produced by the
|
caption |
LaTeX caption, see the |
label |
LaTeX label, see the |
align |
alignment specification, see the |
digits |
number of digits to display, see the |
display |
format of the columns, see the |
... |
additional arguments to be passed to |
LaTeX table representing the confusion matrix
Willem Talloen and Tobias Verbeke
xtable method for summary.mcrPlot objects
## S3 method for class 'summary.mcrPlot' xtable(x, caption = NULL, label = NULL, align = NULL, digits = NULL, display = NULL, ...)## S3 method for class 'summary.mcrPlot' xtable(x, caption = NULL, label = NULL, align = NULL, digits = NULL, display = NULL, ...)
x |
object of class 'summary.mcrPlot' as produced by the
|
caption |
LaTeX caption, see the |
label |
LaTeX label, see the |
align |
alignment specification, see the |
digits |
number of digits to display, see the |
display |
format of the columns, see the |
... |
additional arguments to be passed to |
LaTeX table representing the summary of the mcrPlot output, i.e. the optimal number of features, the mean MCR and the standard deviation on the MCR for each of the classification methods used.
Willem Talloen and Tobias Verbeke
summary.mcrPlot, mcrPlot,
xtable
data(nlcvRF_SS) mp <- mcrPlot(nlcvRF_SS, plot = FALSE) smp <- summary(mp) xtable(smp)data(nlcvRF_SS) mp <- mcrPlot(nlcvRF_SS, plot = FALSE) smp <- summary(mp) xtable(smp)