CVST (version 0.2-3)

constructLearner: Construction of Specific Learners for CVST

Description

These methods construct a CVST.learner object suitable for the CVST method. These objects provide the common interface needed for the CV and fastCV methods. We provide kernel logistic regression, kernel ridge regression, support vector machines and support vector regression as fully functional implementation templates.

Usage

constructLearner(learn, predict)
constructKlogRegLearner()
constructKRRLearner()
constructSVMLearner()
constructSVRLearner()

Arguments

learn

The learning methods which takes a CVST.data and list of parameters and return a model.

predict

The prediction method which takes a model and CVST.data and returns the corresponding predictions.

Value

Returns a learner of type CVST.learner suitable for CV and fastCV.

Details

The nu-SVM and nu-SVR are build on top the corresponding implementations of the kernlab package (see reference). In the list of parameters these implementations expect an entry named kernel, which gives the name of the kernel that should be used, an entry named nu specifying the nu parameter, and an entry named C giving the C parameter for the nu-SVR.

The KRR and KLR also expect kernel and necessary other parameters to construct the kernel. Both methods expect a lambda parameter and KLR additonally a tol and maxiter parameter in the parameter list.

Note that the lambda of KRR/KLR and the C parameter of SVR are scaled by the data set size to allow for comparable results in the fast CV loop.

References

Alexandros Karatzoglou, Alexandros Smola, Kurt Hornik, Achim Zeileis. kernlab - An S4 Package for Kernel Methods in R Journal of Statistical Software Vol. 11, Issue 9, Nov 2004. DOI: 10.18637/jss.v011.i09.

Volker Roth. Probabilistic discriminative kernel classifiers for multi-class problems. In Proceedings of the 23rd DAGM-Symposium on Pattern Recognition, pages 246--253, 2001.

See Also

CV fastCV

Examples

# NOT RUN {
# SVM
ns = noisySine(100)
svm = constructSVMLearner()
p = list(kernel="rbfdot", sigma=100, nu=.1)
m = svm$learn(ns, p)
nsTest = noisySine(1000)
pred = svm$predict(m, nsTest)
sum(pred != nsTest$y) / getN(nsTest)
# Kernel logistic regression
klr = constructKlogRegLearner()
p = list(kernel="rbfdot", sigma=100, lambda=.1/getN(ns), tol=10e-6, maxiter=100)
m = klr$learn(ns, p)
pred = klr$predict(m, nsTest)
sum(pred != nsTest$y) / getN(nsTest)
# SVR
ns = noisySinc(100)
svr = constructSVRLearner()
p = list(kernel="rbfdot", sigma=100, nu=.1, C=1*getN(ns))
m = svr$learn(ns, p)
nsTest = noisySinc(1000)
pred = svr$predict(m, nsTest)
sum((pred - nsTest$y)^2) / getN(nsTest)
# Kernel ridge regression
krr = constructKRRLearner()
p = list(kernel="rbfdot", sigma=100, lambda=.1/getN(ns))
m = krr$learn(ns, p)
pred = krr$predict(m, nsTest)
sum((pred - nsTest$y)^2) / getN(nsTest)
# }