Kernel Enabled K-Means Algorithm
Svn Vishwanathan, Narasimha Murty M · NOT FOUND REPOSITORY (Indian Institute of Science Bangalore) · 2002
We present a novel method to learn arbitrary cluster boundaries by extending the k-means algorithm to use Mercer kernels. We inter- pret each cluster centroid as a linear com- bination of the cluster points in the higher dimensional space and use this formulation to kernel enable the k-means algorithm. The advantage of this formulation is that we work in the higher dimensional kernel space where it is easier to nd smooth surfaces which separate points belonging to di clus- ters. We also extend our formulation to the non separable case by penalizing the violat- ing points quadratically. We show that the clusters obtained vary as a function of the width parameter of the Gaussian kernel.