computeCompositeKernel: computeCompositeKernel

Description

compute a kernel matrix that combines a standard kernel with a user-specified partial labelling. Labellings are handled according to "method" parameter.

Usage

computeCompositeKernel(data, alpha=3, method="neighbors", pgaussian=TRUE)

Arguments

data

n x (d+1) matrix of d-dimensional row-elements. Last matrix column is a label specification in 1..n (NA if element is unlabelled)

alpha

parametrizes the kernel matrix transformation (see Bruneau and Otjacques, 2012)

method

transformation method, being either "simple" or "neighbors" (see Bruneau and Otjacques, 2012)

pgaussian

if TRUE, use the p-gaussian kernel function (see Francois et al. 2005). Else use the classic gaussian kernel, ie p-gaussian function with p=2 and sigma set to the maximal pairwise distance between elements of the data set.

Value

n x n kernel matrix.

References

Bruneau, P. and Otjacques, B. (2012) Including semi-supervision in a kernel matrix, with a view to interactive visual clustering. Tech Report hal-00751407, CRP Gabriel Lippmann. Francois, D., Wertz, V. and Verleysen, M. (2005) About the locality of kernels in high-dimensional spaces. International Symposium on Applied Stochastic Models and Data Analysis. 238-245.

Examples

Run this code

data(iris)
irisdat <- as.matrix(cbind(iris[,1:4], rep(NA, 150)))
irisdat[sample(1:150,4),5] <- 1:4
kernel <- computeCompositeKernel(irisdat)