stam.cv {stam}R Documentation

Cross Validated Training for StAM

Description

Determine classifiers in leaf nodes and weights in inner nodes as well as best graph shrinkage by cross validated model fitting.

Usage

stam.cv(expression.matrix, classifications, 
        chip = "hgu95av2", root = "GO:0008150", 
        beta = NULL, deltas = NULL, ndeltas = 10, 
        results.per.node = FALSE, old.cv = NULL, 
        pamimagefile = NULL, verbose = FALSE)

Arguments

expression.matrix holds the expression levels. It may be of class exprSet or a plain numeric matrix. In the first case exprs is used to extract the expression levels. The matrix is expected to hold one column per sample and one row per probeset.
classifications This character vector must contain one entry per sample identifying the group it belongs to. Alternatively, if expression.matrix is an exprSet, this may be the name of a phenoData variable.
chip the name of the microarray chip. A meta data package is expected to be found holding the needed annotation, namely the links between probesets and Gene Ontology nodes.
root the GO node used as root of the classifier graph. Only successors of this node are considered during construction of the graph.
beta holds class weights used when judging classifier quality. The default is to set class weights to the corresponding prevalence.
deltas numeric vector holding graph shrinkage candidates. Default is to determine ndelta candidates between 0 and the lowest shrinkage level which removes all leaf nodes.
ndeltas number of automatically determined graph shrinkage candidates determined if deltas is not defined.
results.per.node whether results for each node should be returned
old.cv stamCV object used to modify when PAM fits need not to be recomputed. E.g. used when only beta is adapted.
pamimagefile When this parameter is specified stam.cv tries to read this file and extract a stamCV object to be used as old.cv. If the file does not yet exist, PAM fits are stored there after computation.
verbose when set to TRUE reports summary on each leaf training, otherwise shows a progress bar.

Details

stam.cv uses stam.net to generate a classifier graph for the microarray chip at hand. It then fits a PAM classifier for each leaf node only considering the probesets annotated to the node. Afterwards, in each inner node, weights are attributed to each child according to the childs classification performance. Finally, the weights are shrunken such that most of them become zero. In fact, the best shrinkage level is chosen in a cross validation setting.

Classification performance is evaluated using an inverted deviance like measure which uses weights to overstate specificity of a classifier. Weights for nodes are chosen according to this measure and shrunken by an absolute shrinkage level. For each shrinkage candidate cross validated performance results in terms of graph heterogeneity and classification performance are stored.

Value

An object of class stamCV is returned. Use the methods print and plot to extract information about the cross validation.

Author(s)

Claudio Lottaz

See Also

stamCV-class, plot.stamCV, stam.writeHTML

Examples

# load and prepare some data
library(vsn)
library(golubEsets)
data(golubMerge)
golubNorm <- vsn(golubMerge, describe.preprocessing=FALSE)
golubTrain <- golubNorm[,1:38]

# classify into ALL and AML
# (root is chosen to yield results reasonably fast, 
#  consider GO:0008150 (biological process) to obtain 
#  meaningful results)
golubTrain.cv <- stam.cv(golubTrain, "ALL.AML", chip="hu6800",
                         root="GO:0005576", ndeltas=10)

# get further information
print(golubTrain.cv)
plot(golubTrain.cv, delta=0.6)

[Package stam version 1.0.3 Index]