predict.svmpath {svmpath} | R Documentation |
Provide a value for lambda
, and produce the fitted lagrange alpha
values. Provide values for x
, and get fitted function values or
class labels.
predict.svmpath(object, newx, lambda, type = c("function", "class", "alpha", "margin"),...)
object |
fitted svmpath object |
newx |
values of x at which prediction are wanted. This is a
matrix with observations per row |
lambda |
the value of the regularization parameter. Note that
lambda is equivalent to 1/C for the usual parametrization of
a SVM |
type |
type of prediction, with default "function" . For
type="alpha" or type="margin" the newx argument is not required |
... |
Generic compatibility |
This implementation of the SVM uses a parameterization that is slightly
different but equivalent to the usual (Vapnik) SVM. Here
lambda=1/C.
The Lagrange multipliers are related via
αstar = alpha/lambda, where
alphastar is the usual multiplier, and
alpha our multiplier. Note that if alpha=0
, that
observation is right of the elbow; alpha=1
, left of the elbow;
0<alpha<1
on the elbow. The latter two cases are all support
points.
In each case, the desired prediction.
Trevor Hastie
The paper http://www-stat.stanford.edu/~hastie/Papers/svmpath.pdf, as well as the talk http://www-stat.stanford.edu/~hastie/TALKS/svmpathtalk.pdf.
coef.svmpath, svmpath
data(svmpath) attach(balanced.overlap) fit <- svmpath(x,y,trace=TRUE,plot=TRUE) predict(fit, lambda=1,type="alpha") predict(fit, x, lambda=.9) detach(2)