Diffusion Kernels

Imre Risi Kondor

Abstract

I present a generalization of the familiar Gaussian kernel to graphs and a wide range of discrete objects representable by graphs. Having a kernel will allow us to deploy our whole collection of kernel-based algorithms, such as SVMs, Gaussian Processes, Kernel Regression, Kernel Clustering, etc. when dealing with data that naturally falls on such discrete objects, and might not take kindly to being coerced into Rn.

Back to the Main Page