gmum.r (version 0.2.1)

SVM: Create SVM object

Description

Create and train SVM model object.

Usage

SVM(x, ...)
"SVM"(formula, data, ...)
"SVM"(x, y, core = "libsvm", kernel = "linear", prep = "none", transductive.learning = FALSE, transductive.posratio = -1, C = 1, gamma = if (is.vector(x)) 1 else 1/ncol(x), coef0 = 0, degree = 3, class.weights = NULL, example.weights = NULL, cache_size = 100, tol = 0.001, max.iter = -1, verbosity = 4, class.type = "one.versus.all", svm.options = "", ...)

Arguments

x

Training data without labels in one of the following formats: data.frame, data.matrix, SparseM::matrix.csr, Matrix::Matrix, slam::simple_triplet_matrix

...

other arguments not used by this method.

formula

Can be passed with data instead of x, y pair, formula needs to point to lables column, for example: target~.

data

Can be passed instead of x, y pair with formula to mark the labels column, supported formats are: data.frame, data.matrix

y

Labels in one of the followinf formts: factor, vector. Recommended type is factor

core

Support Vector Machine library to use in traning, available are: 'libsvm', 'svmlight'; default: 'libsvm'

kernel

Kernel type as string, available are: 'linear', 'poly', 'rbf', 'sigmoid'; default: 'linear'

• linear: $x'*w$
• poly: $(gamma*x'*w + coef0)^{degree}$
• rbf: $exp(-gamma*|x-w|^2)$
• sigmoid: $tanh(gamma*x'*w + coef0)$

prep

Preprocess method as string, available are: 'none', '2e'; default: 'none'. For more information on 2eSVM see: http://www.sciencedirect.com/science/article/pii/S0957417414004138

transductive.learning

Option got SVM model to deduce missing labels from the dataset, default: FALSE NOTE: this feature is only available with svmlight library, missing labels are marked as 'TR', if none are found and transductive to TRUE, label 0 will be interpreted as missing

transductive.posratio

Fraction of unlabeled examples to be classified into the positive class as float from $[0,1]$, default: the ratio of positive and negative examples in the training data

C

Cost/complexity parameter, default: 1

gamma

Parameter for poly, rbf and sigmoid kernels, default: 1/n_features

coef0

For poly and sigmoid kernels, default: 0

degree

For poly kernel, default: 3

class.weights

Named vector with weight fir each class, default: NULL

example.weights

Vector of the same length as training data with weights for each traning example, default: NULL NOTE: this feature is only supported with svmlight library

cache_size

Cache memory size in MB, default: 100

tol

Tolerance of termination criterion, default: 1e-3

max.iter

Depending on library:

• libsvm: number of iterations after which the training porcess is killed (it can end earlier is desired tolerance is met), default: 1e6
• svmlight: number of iterations after which if there is no progress traning is killed, default: -1 (no limit)

verbosity

How verbose should the process be, as integer from $[1,6]$, default: 4

class.type

Multiclass algorithm type as string, available are: 'one.versus.all', 'one.versus.one'; default: 'one.versus.one'

svm.options

enables to pass all svmlight command lines arguments for more advanced options, for details see http://svmlight.joachims.org/

Value

SVM model object

Examples

## Not run: 
# # train SVM from data in x and labels in y
# svm <- SVM(x, y, core="libsvm", kernel="linear", C=1)
# 
# # train SVM using a dataset with both data and lables and a formula pointing to labels
# formula <- target ~ .
# svm <- SVM(formula, data, core="svmlight", kernel="rbf", gamma=1e3)
# 
# # train a model with 2eSVM algorithm
# data(svm_breast_cancer_dataset)
# ds <- svm.breastcancer.dataset
# svm.2e <- SVM(x=ds[,-1], y=ds[,1], core="libsvm", kernel="linear", prep = "2e", C=10);
# # more at \url{http://r.gmum.net/samples/svm.2e.html}
# 
# # train SVM on a multiclass data set
# data(iris)
# # with "one vs rest" strategy
# svm.ova <- SVM(Species ~ ., data=iris, class.type="one.versus.all", verbosity=0)
# # or with "one vs one" strategy
# svm.ovo <- SVM(x=iris[,1:4], y=iris[,5], class.type="one.versus.one", verbosity=0)
# 
# # we can use svmlights sample weighting feature, suppose we have weights vector
# # with a weight for every sample in the traning data
# weighted.svm <- SVM(formula=y~., data=df, core="svmlight", kernel="rbf", C=1.0,
#                     gamma=0.5, example.weights=weights)
# 
# # svmlight alows us to determine missing labels from a dataset
# # suppose we have a labels y with missing labels marked as zeros
# svm.transduction <- SVM(x, y, transductive.learning=TRUE, core="svmlight")
# 
# # for more in-depth examples visit \url{http://r.gmum.net/getting_started.html}
# ## End(Not run)