Creates and configure all objects needed for a ``variable selection for classificacion'' problem. It configures Gene, Chromosome, Niche, World, Galgo and BigBang objects.
configBB.VarSel(
file=NULL,
data=NULL,
classes=NULL,
train=rep(2/3,333),
test=1-train,
force.train=c(),
force.test=c(),
train.cases=FALSE,
main="project",
classification.method=c("knn","mlhd","svm","nearcent",
"rpart","nnet","ranforest","user"),
classification.test.error=c(0,1),
classification.train.error=c("kfolds","splits","loocv","resubstitution"),
classification.train.Ksets=-1,
classification.train.splitFactor=2/3,
classification.rutines=c("C","R"),
classification.userFitnessFunc=NULL,
scale=(classification.method[1] %in% c("knn","nearcent","mlhd","svm")),
knn.k=3,
knn.l=1,
knn.distance=c("euclidean", "maximum", "manhattan",
"canberra", "binary", "minkowski", "pearson", "kendall", "spearman",
"absolutepearson","absolutekendall", "absolutespearman"),
nearcent.method=c("mean","median"),
svm.kernel=c("radial","polynomial","linear","sigmoid"),
svm.type=c("C-classification", "nu-classification", "one-classification"),
svm.nu=0.5,
svm.degree=4,
svm.cost=1,
nnet.size=2,
nnet.decay=5e-4,
nnet.skip=TRUE,
nnet.rang=0.1,
geneFunc=runifInt,
chromosomeSize=5,
populationSize=-1,
niches=1,
worlds=1,
immigration=c(rep(0,18),.5,1),
mutationsFunc=function(ni) length(ni),
crossoverFunc=function(ni) round(length(ni)/2,0),
crossoverPoints=round(chromosomeSize/2,0),
offspringScaleFactor=1,
offspringMeanFactor=0.85,
offspringPowerFactor=2,
elitism=c(rep(1,9),.5),
goalFitness=0.90,
galgoVerbose=20,
maxGenerations=200,
minGenerations=10,
galgoUserData=NULL,
maxBigBangs=1000,
maxSolutions=1000,
onlySolutions=FALSE,
collectMode="bigbang",
bigbangVerbose=1,
saveFile="?.Rdata",
saveFrequency=50,
saveVariable="bigbang",
callBackFuncGALGO=function(...) 1,
callBackFuncBB=plot,
callEnhancerFunc=function(chr, parent) NULL,
saveGeneBreaks=NULL,
geneNames=NULL,
sampleNames=NULL,
bigbangUserData=NULL
)The file containing the data. First row should be sample names. First column should be variable names (genes). Second row must be the class for every sample if classes is not provided.
If a file is not provided, data is the a data matrix or data frame with samples in columns and genes in rows (with its respective colnames and rownames set). If data is provided, class must be specified.
if a file is not provided, specifies the classes for the data. If the file is provided and classes is specified, the second row of the file is considered as data.
A vector of the proportion of random samples to be used as training sets. The number of sets is determined by the length of train. The train+test should never be greather than 1. All sets are randomly chosen with the same proportion of samples per class than the original sample set.
A vector of the proportion of random samples to be used as testing sets. The number of sets is determined by the length of train. All sets are randomly chosen with the same proportion of samples per class than the original sample set.
A vector with sample indexes forced to be part of all training sets.
A vector with sample indexes forced to be part of all test sets.
If TRUE, the same number of cases for each class. If numeric vector, then it is interpreted as the number of samples in training per class
A string or ID related to your project that will be used in all plots and would help you to distinguish results from different studies.
The method to be used for classification. The current available methods (in this package) are "knn", "mlhd", "svm", "nearcent" (nearest centroid), "rpart" (recursive partitioning trees), and "nnet" (neural networks, experimental, not recommendable), "ranforest" for Random Forest, "user" is for classification problems but the user provides a specific function.
Vector of two weights specifing how the fitness function is evaluated to compute the test error. The first value is the weight of training and the second the weight of test. The default is c(0,1) which consider only test error. The sum of this values should be 1.
Specify how the training set is divided to compute the error in the training set (in evolve method for Galgo object). The fitness function really compute 1-error where error is always computed from the proportion of samples that has been incorrectly classified. "kfolds" (k-fold-cross-validation) compute K non overlapping sets (classification.train.Ksets) attempting to conserve class proportions. "splits" compute K (classification.train.Ksets) random splits. "loocv" (leave-one-out-cross-validation) compute K=training samples. "resubstitution" no folding at all; it is faster and provided for quick overviews.
The number of training set folds/splits. Negative means automatic detection (n=samples, max(min(round(13-n/11),n),3)).
When classification.train.error=="splits", specifies the proportion of samples used in spliting the training set.
For most of the methods, R and C code has been provided. C code is preferred for performance reason, however finding mistakes is easier in R. Besides, the example code could be used as a guide for new user fitness functions. "rpart" has not C code. "svm" has only some improvments removing redundancy checks.
For classification.method == "user", specify the function that would be used to compute the accuracy and class prediction. The required prototype is function(chr, parent, tr, te, result) where chr is the chromosome to be evaluated, a convertion using as.numeric is commonly needed to extract the exact values from the chromosome. parent would be the BigBang object where all their variables are exposed. The fitness function commonly use parent$data$data, which has been trasposed. tr is the vector of samples (rows) that MUST be used as training and te the samples that must be used as test. They can correspond to training and test in the evolution or in any other context (as the computation of the confusion matrix or the forward selection). The fitness function should return the result in two different formats, which is specified in the result parameter. result is 0 (zero) when the predicted class for the test is required (as an integer, not as a factor) otherwise the it is expected the number of correctly classified samples from the test vector.
TRUE instruct to scale all rows for zero mean and unitary variance. By default, scale is TRUE when classification.method is "knn","nearcent","mlhd", or "svm".
For KNN method, knn.k is the number of nearest neighbours to consider.
For KNN method, knn.l is the number of minimum neighbours needed to predict a class.
The distance to be used in KNN method. Possible values are "euclidean", "maximum", "manhattan", "canberra", "binary", "minkowski", "pearson", "kendall", "spearman", "absolutepearson","absolutekendall", "absolutespearman" (see dist method).
For nearest centroid method, nearcent.method specify the method for computing the centroid ("mean", "median").
Parameter passed to nnet.
Parameter passed to nnet.
Parameter passed to nnet.
Parameter passed to nnet.
For SVM (support vector machines) method, specify the kernel method "radial","polynomial","linear" or "sigmoid" (see svm method in e1071 package).
For SVM method, specify the type of classificacion.
For SVM method and nu-classification specify the nu value.
For SVM method and polynomial kernel, specify the degreee value.
For SVM method, specify the C value (cost).
Parameters for neural networks classification. See nnet package.
The function that provides random values for genes. The default is runifInt, which generates a random integer value with a uniform distribution.
Specify the chromosome size (the number of variables/genes to be included in a model). Defaults to 5. See Gene and Chromosome objects.
Specify the number of chromosomes per niche. Defaults is min(20,20+(2000-nrow(data))/400). See Chromosome and Niche objects.
Specify the number of niches. Defaults to 2. See Niche, World and Galgo objects.
Specify the number of worlds. Defaults to 1. See World and Galgo objects.
Specify the migration criteria.
Specify the function that returns the number of mutations to perform in the population.
Specify the function that returns the number of crossover to perform. The default is the length of the niche divided by 2.
Specify the active positions for crossover operator. Defaults to a single point in the middle of the chromosome. See Niche object.
Scale factor for offspring generation. Defaults 1. See Niche object.
Mean factor for offspring generation. Defaults to 0.85. See Niche object.
Power factor for offspring generation. Defaults to 2. See Niche object.
Elitism probability/flag/vector. Defaults to c(1,1,1,1,1,1,1,1,1,0.5) (elitism present for 9 generations followed by a 50% chance, then repeated). See Niche object.
Specify the desired fitness value (fraction of correct classification). Defaults to 0.90. See Galgo object.
verbose parameter for Galgo object.
Maximum number of generations. Defaults to 200. See Galgo object.
Minimum number of generations. Defaults to 10. See Galgo object.
Additional user data for the Galgo object. See Galgo object.
Maximum number of bigbang cycles. Defaults to 1000. See BigBang object.
Maximum number of solutions collected. Defaults to 1000. See BigBang object.
Save only when a solution is reach. Defaults to FALSE (to use all the information, then a filter can be used afterwards). See BigBang object.
information to collect. Defaults to "bigbang". See BigBang object.
Verbose flag for BigBang object. Defaults to 1. See BigBang object.
File name where the data is saved. Defaults to NULL which implies the name is a concatenation of classification.method, method specific parameters, file and ".Rdata". See BigBang object.
How often the ``current'' solutions are saved. Defaults to 50. See BigBang object.
Internal R variable name of the saved file. Defaults to ``bigbang''. See BigBang object.
callBackFunc for Galgo object. See Galgo object.
callBackFunc for BigBang object. See BigBang object.
callEnhancerFunc for BigBang object. See BigBang object.
saveGeneBreaks vector for BigBang object. Defaults to NULL which means to be computed automatically (recommended). See BigBang object.
The gene (variable) names if they differ from the first column in file or rownames(data).
The sample names if they differ from first row in file or colnames(data).
Additional user data for BigBang object (stored in data variable in BigBang object returned).
A ready to use bigbang object.
*** TO DO: EXPLAIN THE STRUCTURE OF "DATA" ***
Wrapper function. Configure all objects from parameters.
# NOT RUN {
bb <- configBB.VarSel(...)
bb
blast(bb)
# }
Run the code above in your browser using DataLab