Low cost estimation of $\sigma$ for SVM using local features
Vanessa Gomez-Verdejo, Miguel Lazaro-Gredilla and Emilio Parrado-Hernandez
Unpublished, Madrid, Spain.
We investigate low cost methods to select the spread parameter in RBF kernels
for Support Vector Machines. These methods try to gain information
about the local structure of the dataset from the performance of
simple local methods such as
k-nearest neighbors. Empirical results in UCI datasets show that the
proposed methods can be used as an alternative to the standard
crossvalidation with the advantage that one does not need to fix a
range of values of the parameter.