svmFeatureSelectionLOOCV: Nested variable selection using LOOCV

In ddgraph: Distinguish direct and indirect interactions with Graphical Modelling

Description

Nested variable selection using LOOCV

Usage

svmFeatureSelectionLOOCV(obj, selectionMode="direct", alpha=0.1, p.value.adjust.method="none",
    test.type="mc-x2", mc.replicates=5000, cost.range=logseq(0.01,
    1e+05, 8), gamma.range=logseq(1e-05, 100, 8), max.prop.SV=0.9,
    kernel="radial", skip.DDGraph=FALSE)

Arguments

obj	the DDDataSet object
selectionMode	which variables to take, possible values: "direct" (alias "p"), "direct and joint" (alias "ps"), "joint if no direct" (alias "snp")
alpha	the alpha cutoff to use
p.value.adjust.method	the p value adjustment for multiple testing to be applied
test.type	the type of conditional independence test to be used
mc.replicates	the number of Monte-Carlo replicates when determining p values
cost.range	the range of cost parameter values to evaluate
gamma.range	the range of gamma parameter values to evaluate
max.prop.SV	the maximal proportion of support vectors to number of data points (rows in d)
kernel	kernel type to use (takes valid package e1071 names like "radial")
skip.DDGraph	if to skip DDGraph-based variable selection

Details

A function to select variables in nested way using the following algorithm:

1. repeat for each row in dataset:

   1. make new DDDataSet by removing one row and apply DDGraphs to select features

   2. select best parameters using recalculateSVMparams (i.e. in an inner LOOCV loop)

   3. make the classifier with best parameters and calculate output on the unseen row (removed in step 1)

2. return the collected predictions from step 1.3

Value

the predictions for class labels from LOOCV

ddgraph documentation built on Nov. 17, 2017, 10:50 a.m.