CL = SVMKERNELLEARN(K, Y) uses LIBSVM to train an SVM with custom kernel K and labels Y. The result CL is a structure with fields:
CL.LIBBSVM_CL : SVM (libsvm format) CL.RBF : Using RBF transformation? (0/1) CL.GAMMA : RBF parameter CL.LABELS : Category labels
Options:
Type ['C']
Set the SVM type to 'C' or 'nu'.
C [1]
Set the SVM C parameter. The C parameter establishes the trade-off between maximizing the margin of the decision function from the correctly classified data and the number of misclassified data. A large C gives more importance to reducing the number of mistakes, but may increase overfitting.
Nu [.5]
Set the nu-SVM nu parameter.
RBF [0]
Enable RBF transformation. Assuming that the input argument K is actually a metric, the kernel is defined as K' = EXP(- gamma K).
Gamma [[]]
GAMMA constant of the RBF transformation.
Balance [0]
Enable data balancing. Balancing reweights the data so that the empirical error term (see C option) in the SVM cost functional is computed assuming that the labels are equally probable. Balancing affects the value of the C parameter for each sample, increasing its value for the less represented labels.
CrossValidation [0]
Perform N-fold cross validation to determine the optimal value of the paramter C. In this case specifying C has no effect.
Verbosity [0]
Set verbosity level.
Debug [0]
Print debugging informations.
Probability [0]
Return probability of classification instead of the decision value.
LIMITATIONS. Currently, only C-SVM is supported.
See also SVMKERNELTEST().