liquidSVM (version 1.2.2)

liquidSVM-package: liquidSVM for R

Description

Support vector machines (SVMs) and related kernel-based learning algorithms are a well-known class of machine learning algorithms, for non-parametric classification and regression. liquidSVM is an implementation of SVMs whose key features are:

• fully integrated hyper-parameter selection,

• extreme speed on both small and large data sets,

• full flexibility for experts, and

• inclusion of a variety of different learning scenarios:

  • multi-class classification, ROC, and Neyman-Pearson learning, and

  • least-squares, quantile, and expectile regression

Further information is available in the following vignettes:

demo	liquidSVM Demo (source, pdf)
documentation	liquidSVM Documentation (source, pdf)

Known issues

Interruption (Ctrl-C) of running train/select/test phases is honored, but can leave the C++ library in an inconsistent state, so that it is better to save your work and restart your R session.

liquidSVM is multi-threaded and is difficult to be multi-threaded externally, see documentation

Details

In liquidSVM an application cycle is divided into a training phase, in which various SVM models are created and validated, a selection phase, in which the SVM models that best satisfy a certain criterion are selected, and a test phase, in which the selected models are applied to test data. These three phases are based upon several components, which can be freely combined using different components: solvers, hyper-parameter selection, working sets. All of these can be configured (see Configuration) a

For instance multi-class classification with \(k\) labels has to be delegated to several binary classifications called tasks either using all-vs-all (\(k(k-1)/2\) tasks on the corresponding subsets) or one-vs-all (\(k\) tasks on the full data set). Every task can be split into cells in order to handle larger data sets (for example \(>10000\) samples). Now for every task and every cell, several folds are created to enable cross-validated hyper-parameter selection.

The following learning scenarios can be used out of the box:

mcSVM

binary and multi-class classification

lsSVM

least squares regression

nplSVM

Neyman-Pearson learning to classify with a specified rate on one type of error

rocSVM

Receivert Operating Characteristic (ROC) curve to solve multiple weighted binary classification problems.

qtSVM

quantile regression

exSVM

expectile regression

bsSVM

bootstrapping

To calculate kernel matrices as used by the SVM we also provide for convenience the function kern.

liquidSVM can benefit heavily from native compilation, hence we recommend to (re-)install it using the information provided in the installation section of the documentation vignette.

References

http://www.isa.uni-stuttgart.de

See Also

init.liquidSVM, trainSVMs, predict.liquidSVM, clean.liquidSVM, and test.liquidSVM, Configuration;

Examples

# NOT RUN {
set.seed(123)
## Multiclass classification
modelIris <- svm(Species ~ ., iris)
y <- predict(modelIris, iris)

## Least Squares
modelTrees <- svm(Height ~ Girth + Volume, trees)
y <- predict(modelTrees, trees)
plot(trees$Height, y)
test(modelTrees, trees)

## Quantile regression
modelTrees <- qtSVM(Height ~ Girth + Volume, trees, scale=TRUE)
y <- predict(modelTrees, trees)

## ROC curve
modelWarpbreaks <- rocSVM(wool ~ ., warpbreaks, scale=TRUE)
y <- test(modelWarpbreaks, warpbreaks)
plotROC(y,warpbreaks$wool)
# }